I don’t do that anymore.

OR:ON, the new monitoring product we’ve been building at Hexact, has an MCP server. In plain words, that means Claude (ChatGPT or any AI tool that supports MCP) can read my watches and every change they’ve caught. So now I just ask. And more and more, I don’t even ask a question. I ask for a report.

That shift changed how I use monitoring. I decide what goes in the report, what I care about this week, and who it’s for.

Here’s what that looks like in practice.

Your monitoring data, where you already work

Most monitoring tools keep your data locked in their own app. You get an email, you click through, you read a diff, you close the tab. If you want a summary for your team, you write it yourself.

With OR:ON connected to AI tool, the data comes to you instead. Any watch and any change it caught is something you can ask about, group or reshape. You can also mix it with whatever else Claude has access to: your notes, your CRM, your sales numbers, or the Second Brain by Hexact.

The question stops being “what changed?” and becomes “what should I do about it?” AI can answer the second one because it sees the whole picture at once.

None of this works if the page won’t load

A report is only as good as the data behind it. And getting the data is the hard part. Many sites block automated visitors, hide content behind heavy JavaScript, or show different pages by country. Most monitoring tools quietly fail on those pages, and you never find out what you missed.

Today OR:ON loads 98.1% of the pages our users ask it to watch. That’s far ahead of anything else on the market, and we’re still working on the last 1.9%.

Don’t take my word for it. Take a page that no other monitoring tool could open for you and paste it into OR:ON’s preview. There’s a good chance you’ll see it load for the first time.

Two examples from my week

Tracking a site under heavy development. Valedo.app ships changes almost every day. New pages and layout changes go live, often from several people at once. OR:ON watches the sitemap, the key pages and the tech stack, so every change gets logged, whether anyone announced it or not. Last week that was 197 new pages plus a steady stream of edits to existing ones. Every Friday I ask Claude: “What changed on Valedo this week, and did anything break or disappear?” I get a clean changelog the team never had to write, and twice now it caught a section that went missing after a deploy before any user noticed.

A market check. For Grillyan, OR:ON watches three competitor review pages, their service area pages, and the stock pages of a parts supplier. The alerts only fire on what matters, like a price or a new county. Every Monday I ask for a short market check. Last week’s:

One competitor cut their deep clean price from by $99. Another added Palm Beach County. Igniter kits for two of our top brands are back in stock.

Three lines, three decisions. Before, that information lived in separate tabs, and I only saw it if I went looking.

Same data, any format

This is the part I underestimated. Once the data is in Claude, the format is up to you. A few prompts I’ve used in the last two weeks:

• “Turn this week’s OR:ON changes into a one-page PDF for my Monday meeting.”

• “Make a chart of new Valedo pages per day for the last month.”

• “Write a Slack update for the team. Three bullets, no fluff.”

• “Export every competitor price change since September as a CSV.”

• “Compare this week to last week. What’s new, what stopped changing?”

Also new: a Second Brain update

We shipped a new version of Second Brain this week. Three things changed:

• Faster imports. Big archives and large CSV files now load noticeably quicker.

• UI improvements. Cleaner screens and fewer clicks.

• Bug fixes. A long list of small ones, mostly around imports.

Try it this week

1. Set up one watch in OR:ON on a page you check by hand today. A competitor’s pricing page is a good start.

2. Connect OR:ON to your favorite AI tool and ask: “Give me a one-page report of everything that changed this week, ranked by what I should act on first.”

3. Update or install Second Brain.

But a lot of what matters never shows up on a page you’d think to open. It shows up in a feed, a data file, a domain record, a script in the code, or a server that stops answering. OR:ON watches those too, and we’ve barely mentioned them.

Every Watch works the same way. You add the address, preview the first result, and pick how often it checks and on which days. Later checks compare against that first result. You choose what counts as an alert: any change, something new, something removed, or a sentence you write. Alerts go to email, Slack, Discord, Teams or Telegram.

What changes is the question each Watch answers. Here are seven, with one example each.

1. What did they publish?

RSS Watch. Alerts when an item in a feed is new or changed, with the title and the link.

Most people think of RSS as a way to follow blogs. The more useful feeds come from places that publish things you have to act on.

Take the tools your business runs on. Most software projects on GitHub publish their releases as a feed at github.com/owner/project/releases.atom. Point an RSS Watch at the plugin, library or app your site depends on, and you hear about a new version, or a security fix, the day it ships. You don’t have to wait for something to break first.

The same goes for government agencies, standards bodies and industry groups. Many of them publish notices, recalls and rule changes as feeds that nobody reads until a customer asks about them.

Available on every plan, including Free.

2. What are they building, and what did they drop?

Sitemap Watch. Alerts when a page address is added to a sitemap or removed from it.

A sitemap is the list of pages a site wants search engines to find. Companies update it with every launch, often before the new page is linked from anywhere.

A competitor’s sitemap gaining /integrations/salesforce or /pricing/enterprise tells you where they’re heading. The removals tell you even more. When a whole product line disappears from a sitemap, someone made a decision, and you’ll know about it before any announcement.

Now do the same for your own site. A redesign or a plugin update can quietly drop hundreds of addresses from your sitemap. Search engines use that list to find and recheck pages, so the missing ones get found and updated less often. A Sitemap Watch on your own domain tells you the same day, while it’s still an easy fix.

Available on every plan, including Free.

3. Did the data I depend on change?

JSON Watch and API Watch. Both alert when a field is added, removed, or its value changes.

A lot of data sits one step behind the page. Many Shopify stores, for example, publish their whole catalog at /products.json: every product, variant, price and stock flag. A JSON Watch on that address catches a price change or a product going out of stock across the entire store, not just the one page you happened to bookmark.

The API Watch is for data you can’t get with a plain link. You set the method, headers and body, plus the status you expect back. That covers the services your business runs on: shipping rates, supplier inventory, a payment provider’s response. When a vendor renames a field or changes a value format, your integration breaks and nobody sends you an email. The API Watch does.

JSON Watch: every plan, including Free. API Watch: Plus and Scale.

4. Is my site actually up?

Uptime Watch. Alerts when a host goes down and again when it comes back, with the status and how long it was down.

Most owners find out their site was down from a customer. The checkout stopped answering at 2 AM, and the first sign was an angry email at 9.

An Uptime Watch tells you at 2 AM, in the channel your team actually reads. The second alert matters just as much: you learn when it came back and how long it was gone. That’s the number you need when you talk to your host, or explain a bad morning to your boss.

Watch the hosts you don’t own too. If your store depends on a payment provider or a shipping API, an Uptime Watch on that host tells you the problem is theirs before you spend an hour searching your own code.

Plus and Scale.

5. Who just switched tools?

Tech stack Watch. Alerts when a site adds a technology, removes one, or changes its version.

When a company adds a chat tool, moves to a new store platform, or drops an analytics product, it’s spending money and making decisions. That’s a sales signal, and it’s public.

If you sell to companies on a specific platform, put your prospect list into Tech stack Watches and wait for the platform to appear. If you compete with a known product, watch its customers for the opposite. A script that disappears from a customer’s site often means the contract ended and they’re shopping.

Agencies can point it at client sites. A version change you didn’t make, or an old version that never gets updated, is a reason to call the client before something breaks.

Personal, Plus and Scale.

6. Is my domain safe?

WHOIS Watch. Alerts when a domain’s registrar, expiry date, nameservers, status or DNS records change.

These records almost never change. That’s exactly why a change is worth an alert.

Start with your own domain. A new nameserver or registrar you didn’t set up is an early sign of a hijack. A changed MX record means your email is now going somewhere else. Either one can cost you a week of lost mail or a lost domain, and the WHOIS Watch catches it on the next check. Send this one to Slack, not a daily digest.

It also works for a domain you want to buy. For most generic domains like .com, an expired name goes through a renewal grace period, then about 30 days of redemption and 5 days of pending delete before anyone can register it. A WHOIS Watch on its status tells you when it enters each stage, so you’re ready on the day it drops.

Personal, Plus and Scale.

7. What changed under the hood?

Source code Watch. Alerts when a page’s HTML changes, and shows the lines that changed.

Some of the most important changes are invisible on the page.

On your own site, the classic one is a single line: <meta name="robots" content="noindex">. It slips onto a live page after a deploy, the page looks perfect, and search engines quietly drop it. Nobody notices until traffic falls weeks later. A Source code Watch on your key pages shows you that line the day it appears. The same goes for a tracking script that vanished after an update, or a plugin that injected code you never approved.

On someone else’s site, the code tells you what they’re testing. A new A/B testing script, a new ad pixel or a changed page title usually comes before a visible change.

Plus and Scale.

Start with one you can set up tonight

Three of these work on the Free plan: RSS, Sitemap and JSON. Pick the one closest to a question you already have.

• Do it yourself: open orion.hexowatch.com then click add URL and paste your Sitemap, and you’re done. No card needed.

If you used Hexowatch, you know the routine. An email says the page changed. You open two screenshots side by side. You squint. Nothing you care about is different. You close it. After the tenth one, you stop opening them.

It was one of the most common complaints we got for five years.

What actually set off the alerts

We wrote about this so often that “Why am I getting false positives?” had its own section in our FAQ. Here’s the honest list of what triggered most of them:

• Timestamps. “Updated 4 minutes ago” or a date in the footer. Different on every check, so every check was a “change.”

• Rotating banners and ads. A new ad loads, the hero slider moves to slide three. The visual monitor sees new pixels and fires.

• Session elements. A cart counter, a “Welcome back” message, a random A/B test variant. Each visit looked a little different from the last.

• Tiny CSS shifts. A font loads a fraction of a second late and pushes the layout down two pixels. To a screenshot diff, that’s a change across the whole page.

• Half-loaded pages. The check took its snapshot before the page finished loading. Half the content was missing, so we reported a change. Next check it loaded fully, and we reported a change again.

• Block pages. A site served a Cloudflare challenge instead of the real page. Hexowatch compared the challenge to yesterday’s page and called it a change.

• Our browser saw a different page than you did. Old sites, and sites that don’t handle different screen sizes, often rendered one way for us and another way for you.

Then there was the opposite problem. HTML element monitors were precise, so we recommended them for prices. But they were tied to a spot in the page code. When a site rebuilt its markup, the price could stay exactly the same and the monitor would still break or fire. Our own FAQ said it plainly: that’s not a bug, that’s how the web works.

We were right about the web. We were wrong to stop there.

What we told you to do about it

Our advice was always some version of “set it up better”:

• Combine Visual and HTML monitors so design changes don’t trigger alerts.

• Pick stable elements. Exclude the dynamic parts of the page.

• Check less often. Five pages every 15 minutes is 480 visits a day, enough to get blocked on a small site.

• Book a Concierge session and we’ll tune it until false positives are close to zero.

All of that worked, sometimes, but it meant every useful monitor needed someone who understood selectors, delays, and regions to build it, then fix it when the site changed. Most people don’t have that person. So they did the reasonable thing and muted the alerts.

In January we finished months of backend work: smarter page loading, better diff detection, more retries. Success rates went up and false positives dropped. They didn’t go away, because the core problem was still there.

The core problem

Hexowatch compared two versions of a page, as code or as pixels, and reported the difference. It had no idea what the difference meant. To Hexowatch, a new timestamp and a new price were the same thing: some characters changed somewhere on the page.

Every fix we shipped was a rule. Ignore this area. Track that element. Wait three seconds longer. Rules work until the site changes, and sites change all the time.

There’s a name for what this does to people. In hospitals it’s called alarm fatigue. Reviews of hospital alarm studies found that 72% to 99% of clinical alarms are false. Johns Hopkins counted about 350 alarms per patient per day. Nurses react the way you’d expect: they turn the volume down, widen the limits, or tune them out, and the real alarm gets ignored with the rest. A monitoring tool with a bad alert ratio does the same thing to you, with lower stakes.

How OR:ON handles each one

OR:ON starts from what you said matters, then looks at the page. That changes how each of those causes plays out:

• Timestamps, banners, ads, session elements. You asked about the price or the job list. A new ad or a new date is neither, so you don’t hear about it.

• Markup redesigns. OR:ON looks for the price, not a path in the code. If the site rebuilds its layout and the price stays the same, there’s nothing to report.

• Cookie notices. It can close them before taking the snapshot, so a consent popup never ends up in the comparison.

• Block pages and failed loads. A page Watch can retry through a residential proxy instead of reporting a broken page as a change. For stubborn sites, use the proxy on every check.

• Changes you only care about in one direction. Limit a Watch to something new or something removed, and the other direction stays quiet.

When an alert does come, it says what changed in plain words. You don’t open two screenshots to find out it was a banner.

What it still won’t fix

It’s not perfect, and we’d rather say so now.

If you describe what matters vaguely, the alerts will be vague too: “tell me about changes” gets you far more email than “tell me if the Pro plan price changes.” And sites keep finding new ways to block automated visitors. We’ll keep catching up.

Bring back one monitor you muted

You probably have a Hexowatch monitor you stopped reading because it cried wolf. Start there.

• Do it yourself: open orion.hexowatch.com, paste the same URL, and write one sentence about what you actually wanted to know (sign in with your Hexact account).

Back at the last few months of this newsletter, I wrote about AI, small teams, Second Brain, my book, scraping. Almost nothing about website monitoring, the thing many of you signed up for in the first place.

There was a reason. We were rebuilding something new from scratch, and I didn’t want to write about it until I could send you a link and say “go try it.”

Today is the day. OR:ON is live.

From “something changed” to “this changed to that”

Let’s be honest about Hexowatch, for five years it mostly told you one thing: something on this page changed. Then you opened two screenshots and tried to find it yourself. Sometimes it was a price. Often it was a banner that moved, or worst - just half-loaded page.

OR:ON finds it for you and tells you in plain words. Your competitor drops a price, and the alert says “Pro plan price changed from $49 to $39.” A new role shows up on a job board, and you get “New job posted: Senior Product Marketing Manager, Miami.”

You’ll see the before and after next to each other if you want to double check. And OR:ON only writes about the part you asked it to watch. If a popup appears somewhere else on the page, you won’t hear about it.

Ask your favorite AI about it

Every change OR:ON finds is saved. That history turned out to be more useful than we expected.

If you use Claude or ChatGPT, you can connect OR:ON to it and just ask. Something like “What did my competitors change on their pricing this quarter?” Your AI answers from your actual Watches, not from guesses. I use this more than I thought I would.

You can also ask right inside OR:ON. Every answer links to the changes it’s based on, so you can check it yourself.

A normal Tuesday

Your competitor lowers the price on its Pro plan early in the morning. No email, no post on LinkedIn, they just change the number.

Most of the time you’d learn about it weeks later, usually from a customer on a renewal call. Not a fun way to find out.

With OR:ON watching that page, the alert is waiting when you open your email. By lunch you know about the price and you’ve had time to decide if you care.

Setup is one sentence

Paste a link. OR:ON opens the page and shows it to you. Then type what you care about: “Tell me when the price drops below $500.” Or just click the price on the page.

That’s the setup. OR:ON reads the page, works out what you meant and builds the check. On pages it hasn’t seen before, it suggests what’s worth watching.

The pages that don’t want to be watched

Some sites put up a fight. They block automated visitors, hide content behind cookie banners and load differently every time. Hexowatch hit those walls often.

OR:ON gets through. Its success rate at opening pages is more than 10 times higher than Hexowatch’s. It visits pages through residential proxies in the country you choose and closes cookie banners before it looks at anything.

Those sites get smarter. So do we. :)

Why it took us this long

For more than five years we read your support tickets, your feature requests and, yes, your bad reviews. Most of them came down to one wish: just tell me what changed.

A few years ago, we couldn’t build that. A machine looking at a web page saw only code and pixels. It couldn’t tell a price from a phone number or a job post from a footer.

Today’s AI reads a page much like you do. It sees a price as a price. That shift, plus years of your feedback telling us exactly what to fix, is what made OR:ON possible. So we went quiet and built it.

What will you watch first?

Start with a page you already check by hand. A few ideas:

• A competitor’s pricing page. Tell OR:ON to let you know if any plan price changes.

• A product that keeps selling out. Ask it to tell you when it’s back in stock.

• A careers page. Useful if you’re hiring, job hunting or just curious who a competitor is hiring.

• Terms and refund policies. Yours or someone else’s.

• Tenders and grants. They show up with no warning and short deadlines.

• An RSS feed. Point it at a competitor’s blog and hear about every new post without subscribing.

• A sitemap. You’ll see new pages the day they go live, often before they’re linked anywhere.

This is day one. Pages behind a login aren’t supported yet, and the list of what’s coming next is long. Most of it came from you.

Your first Watch takes a minute

Think of one page you still check by hand. Open orion.hexowatch.com, paste the link and tell OR:ON in one sentence what matters. Then close the tab and get on with your day.

The Free plan needs no card, and you sign in with your existing Hexact account.

We’ve been quiet about monitoring for a long time.

I’ve written about Second Brain (Hexact) a lot and mentioned my local service business along the way. I’ve never written about how one led to the other.

Why a service business

I quickly realized I couldn’t rely on a software startup alone. I needed something on the ground, next to Hexact.

Software is still my main business. But with AI it moves faster every month. What was relevant a year ago can be irrelevant today. You move on to the next product while the last one is still running. Some you keep supporting. Some you shut down or sell, and deal with the fallout: unhappy users, refund requests, bad reviews. And what you built last year can be copied in a weekend.

The things I had done before and was comfortable doing, marketing, consulting, agency work, AI is already taking over most of them. A local service business looked like it would outlast all of them. AI can write your ads and answer your phone. It can’t fix a leaking pipe or repair an AC unit.

My business partner and I tried several things first. None worked. “A service business” isn’t a plan. We needed a specific service, in a specific place.

And we knew almost nothing about the market here. We didn’t even know the terms. Home services, field service, skilled trades. We learned them later.

So we started with what I know best. Data.

Hours to collect, months to analyze

I’ve built scraping tools for years. With Hexomatic, pulling local business data across South Florida took hours: listings, categories, reviews, competitors, locations.

We had one question: which niche isn’t overcrowded but still has real potential?

The obvious ones came first. Plumbing, AC repair, roofing, pest control, pool cleaning. Plenty of demand, and plenty of companies already fighting over it. Getting in meant competing on price from day one.

Going deeper is where it got hard. The data sat in spreadsheets and CSV files that didn’t connect. Every question meant opening several files, filtering, comparing, and starting over for the next one.

Collecting took hours. Analyzing took months.

AI didn’t help much

I used AI heavily during that time. I uploaded files, pasted data, asked the same questions in different ways. It worked for small tasks like cleaning a list or summarizing a trend. For the questions that needed all the data at once, it failed. Answers changed between tries. Numbers didn’t add up. I checked most of it by hand.

For that kind of work, not much has changed since.

What the data showed

COVID changed consumer behavior fast. Some shifts were obvious. Online shopping grew as much in a few months as it had in the previous five years. Food delivery and remote work followed, and large companies moved in quickly.

We were looking for something else: changes that created new local demand that service businesses hadn’t caught yet.

So we ran a second scrape, this time of articles about COVID’s impact on business and consumer spending, and added them to the same dataset as the local business data. That let us compare what was changing with what local companies were offering.

One pattern kept appearing. People spent more time at home and invested in their houses, especially their backyards. Most homeowners said they were more interested in upgrading outdoor spaces than before the pandemic, and outdoor kitchens were near the top of the list. In Florida, where people cook outside most of the year, this was even more visible.

Grill manufacturers confirmed it. Traeger and Weber reported record growth, and both went public. Americans bought over 21 million grills and smokers in about two years.

All that equipment was bought in a short window, so it would need cleaning, maintenance, and repair at around the same time. In our local data, very few companies specialized in it.

The niche had growing demand and limited competition. We chose grill cleaning and maintenance in Miami.

My partner and I launched Grillyan in the summer of 2024. The first order came on August 27, my birthday. The first job took the two of us six hours. Today one technician does it in about two and a half.

Grillyan now has five technicians, more than a thousand clients, and is expanding into other counties.

The tool I built for myself

The research taught me one thing. Collecting data is easy now. Making sense of it is the hard part.

So while Grillyan was starting, I built a system for myself. A local database on my computer with everything in one place: Grillyan’s clients and jobs, Hexact’s customers and payments, competitor data, my articles, meeting notes. I connected it to Claude so I could ask questions in plain English and get answers from real records.

Questions that took an afternoon started taking a minute. Claude stopped guessing because it was reading my actual numbers.

For a long time it was only mine.

The moment it clicked

Then I finished my book, Bones: 21 Clichés Running and Ruining Your Business. 211 pages that needed a serious proofread. English isn’t my native language, and over that many pages, a lot of issues pile up.

I gave it to Claude. It did solid work on the first 20 or 30 pages. Then it lost track of tone decisions from earlier chapters, hit its output limit, and every “continue” made it worse. Maybe 15% of the book got a proper review.

So I imported the book into my database instead. Each page became one row of clean text. Now I could run checks across the whole book, like “list every sentence longer than 35 words,” and the quality stayed the same from page 1 to page 211.

That’s when I saw it wasn’t my problem only. Anyone with a long document, a big contact list, or years of business data hits the same wall. AI has the skills. It can’t handle the input in the shape most people give it.

When I wrote about my setup, people replied asking if they could pay me to set it up for them. That became Second Brain.

Big companies, same problem

Right now we’re in active talks with a few very large companies that want to implement this inside their operations. Nothing is signed yet, but the interest is serious.

These companies have invested heavily in technology and AI over the last few years. They still haven’t solved a basic problem: asking a question across a huge amount of their own data and knowing the answer is based on all of it. Not most of it. All of it.

Current AI models can take in a book’s worth of text in one prompt. That doesn’t solve it. The more text you give a model, the more it skims. Some records get read, some get skipped, and you can’t see which.

In coding, you can live with that. Ask AI to review code, it finds some issues. Run it again, it finds new ones it should have caught the first time. Run it a third time, more gaps. You keep iterating until it works.

With numbers and legal documents, you can’t work that way. If a model reviews a thousand contracts and misses forty, or totals payments and skips a few hundred rows, the answer isn’t slightly off. It’s wrong, but it looks right.

Second Brain changes what the AI reads. Claude doesn’t skim raw text. It queries organized records, and a query returns every row that matches. Nothing gets skipped, and nothing depends on which run you’re on.

What’s new this week

The latest Second Brain update adds more file formats and bigger file support, faster imports, bug fixes, and UI improvements.

Download it at brain.hexact.io. New to Second Brain? Start the 7-day free trial, no credit card required.

Want to see it on your own data? Book a demo.

No time to collect and import it yourself? Our Done For You service gathers your data and loads it into Second Brain for you.

That is not a limit of the AI. It is a limit of the inventory.

Your history is not your position

Your own data answers questions about your history. Who paid and went quiet. What you promised in a contract two years ago. Who you met last quarter and never followed up with.

But internal data only tells you your side of the story. It cannot tell you where you stand in the market.

Revenue up eleven percent is a fact about you. Whether eleven percent is good depends on numbers that live somewhere else. You know every customer you have. You have no idea how many you could have, because that number depends on operators you have never heard of.

The data behind those answers is public. Almost nobody has it in a form they can query.

What happens when both halves sit together

Second Brain is a database, not a folder of files, so your internal data and outside data can be queried together.

A list of every company in your category tells you who exists. Put it next to your contacts and your payment history and it tells you which of them already buy from you, which ones you emailed last year and never followed up on, which ones bounced so you stop paying to mail them, and which ones resemble your ten best accounts rather than your ten loudest ones.

A price file from six competitor sites is a spreadsheet. Next to your own transactions it tells you whether the accounts you discounted last year were under real pressure or just asked loudest.

None of that is in the collected file. None of it is in your CRM.

Why the second half stays empty

Collecting outside data properly is its own project.

Finding which sources actually carry the field you need. Getting the extraction to survive a layout change. Handling the same company appearing four times under three spellings. Naming columns so the file still makes sense in November. Then doing it again in six months, because a market snapshot from March describes a market that no longer exists.

Estimated as an afternoon. Abandoned in week two. So the dataset never gets built, and the Second Brain stays half empty.

Say what you want it to know

This is the one part nobody can do for you.

Most requests arrive as a subject. “Companies in my industry.” “My competitors.” “Leads.” Those are wishes. They do not say which geography, which fields, or what happens once the data lands.

A brief that works names a field you can put in a column and a decision you will make with it.

Not “everything about my competitors” but the price of these twelve products on these six sites.

Not “podcasts about my topic” but episodes from these shows, with guests, dates and descriptions.

“Leads” becomes every operator in these three counties with a website and a phone number, matched against my own customer list.

Three questions tell you whether your brief holds up. Does the source publish the actual field, not just the topic. What decision changes when the answer arrives. And what will you match it against.

We go and get it

This is why we built Done For You.

Tell us what you want your Second Brain to know. We find the public sources, collect the data, clean it, and deliver it ready to query.

You never touch a scraping tool. It is a one-time project, not a subscription.

Your private data stays on your machine. We collect the outside data. Second Brain is where the two meet.

See how Done For You works

Write the sentence

Write down the question about your market you cannot answer today. One sentence. Then add the geography, the fields, and what you would do with the answer.

Send it to us at admin@hexact.io, or book a call and we will write the brief together. If it is collectable we tell you how before you pay anything. If it is not, we tell you that too, and there is usually a version of the question that works.

No Second Brain yet? The free trial is enough to load the first dataset. Load one outside source, then ask which of those companies are already in your contacts.

The data isn’t gone. It’s sitting somewhere you can’t really use anymore.

Here are five setups from emails I’ve answered recently, with details changed so nobody’s identifiable.

The export that stops rotting

Many users now send their Hexomatic workflow results straight into their Second Brain database instead of downloading another CSV.

Same data, completely different shelf life. Refreshed on a schedule, it’s still answering questions in November. And sitting next to your contacts it answers better ones than it could alone. Which of these have I already emailed. Which are already paying me. Which did I talk to last year and never follow up.

If you have a Hexomatic account you already have the raw material for this. It’s usually the fastest first import.

Half a million rows nobody could open

One user had a company database of over half a million rows. Size, financials, category, contacts, websites. The spreadsheet took a minute to open and froze on anything real.

He owned that data and couldn’t touch it.

Uploading it to an AI doesn’t solve it either. Half a million rows won’t fit in a context window, so the AI ends up working with pieces of the dataset instead of querying the whole thing.

As a table it’s a different object. Filter by size and revenue in one pass. Pull every company in a category with no website. Then match the whole list against his own contacts, so his team stops cold-emailing people who are already clients.

Thirty gigabytes, no system

Another one wanted nothing to do with connectors. No Notion, no Stripe, no calendar. Just a folder on his laptop, about thirty gigs, years of contracts and proposals piled up. Can I work with this?

Easiest setup there is. Point it at the top folder and it walks everything underneath. PDF, DOCX, TXT, MD, RTF.

Files come in as indexed sections rather than whole documents. Once it’s indexed, the AI can search across thirty gigs without loading thirty gigs into the conversation. Ask about one clause from 2022 and you get the paragraphs back, not the file.

He hadn’t thought about the comparison side. Which agreements still carry a clause you stopped offering. Which proposals promised something you don’t do anymore. By hand that’s a week, so you won’t do it.

What this saves in AI context

When you hand an AI large files directly, the content has to be processed and pushed into the model’s context. Start a new session and much of that work happens again.

Second Brain indexes the material once, at import. After that a query retrieves the six sections that matter instead of two hundred pages. You pay for the reading once instead of every session, and the gap widens as your archive grows.

Five years of research

One setup has nothing to do with customers.

A user is studying a language and had the same problem in a different shape. Vocabulary in a spreadsheet, phrases in a notebook, a few apps that each track part of it. He wanted an honest count of what he actually knew and how it was moving.

We ended up building the Learn section because enough people described that exact thing. The same idea applies to any long project you’ve been accumulating for years. Papers, notes, drafts, interviews. Finding one file was never the hard part. Asking a question across all of them is.

Not betting on one AI vendor

Someone asked whether he could use it with local models instead of depending on one hosted AI company.

That’s the point of keeping the database local and connecting over MCP. Your data sits on your machine and compatible tools query it. You’re not building your knowledge base inside one vendor’s account, which also means it doesn’t depend on us being here.

Start with one pile

Don’t organize everything first. That’s another project you’ll never start.

Nobody above did that. They each picked the thing that had been bothering them and imported that one thing.

The export from March. The spreadsheet that won’t open. The folder of contracts.

Download at brain.hexact.io. Seven days free trial, no card required.

Import one source, then ask it something you couldn’t answer this morning. If you import a customer export, start here: which customers paid me last year and have gone quiet this year.

Want help setting it up?

Book a call and we’ll go through it together on your machine. Your imports, your order, and the questions worth asking once it’s in. Nothing leaves your computer, including during the call.

Here is the part people miss. Google does not just show a page’s meta description. It builds the snippet from the part of the page that matches your query. Google rewrites the author-written description in roughly 60 percent of results. Which means your keyword decides what text comes back.

Search “Nike Air Max 90” and the snippet describes the shoe. Search “Nike Air Max 90 price” and the snippet tends to contain a dollar figure. Same page, different extraction, because you told Google which part of the page you cared about.

That flips the workflow. Instead of scraping a source to get a phone number, you write a keyword that forces the phone number into the snippet, then scrape the search results.

One keyword is a search. Four hundred keywords is a database.

This is the part that makes it worth doing, and it is the part most people skip.

Running one query in Google and copying the answer is not automation. The unlock is that a keyword is a template with variables in it, and every variable has a list behind it. Geography. Job title. Product model. Company name. Zip code. School district. Multiply the lists together and one afternoon of setup produces thousands of rows.

[niche] [city] phone number is not one keyword. With 40 niches and 25 cities it is 1,000 keywords. At 10 results each that is 10,000 rows, and every row has a position, a title, a description and a URL.

In Hexomatic you paste all of them into Data input, one per line, and add the Google Search automation once. The workflow does not care whether the list has 5 rows or 5,000. That is the whole difference between searching and extracting.

Two things make the batch usable instead of messy:

Keep the pattern identical across the batch. Same word order, same field words, same operators, only the variable changes. Google generates snippets per query, so consistent queries return consistent snippet shapes, and consistent shapes are what let you parse 10,000 rows with one rule instead of ten.

The 10 hacks

Each one below is a pattern. The bracketed part is your list.

1. Public sector contacts without building a directory scraper

School districts, municipalities and universities publish staff contacts on hundreds of small pages that all look different. Building one recipe per district is a waste.

Pattern: [district name] school principal phone email, [county] school district superintendent contact phone, site:.k12.fl.us [district] principal email

Multiply by: every district or county in your target states. Florida alone gives you 67 counties. Add three phrasings and you are running 201 keywords for full coverage of one state.

2. Email addresses at a specific company

Pattern: "@[domain]" email, site:[domain] "@[domain]", [domain] contact email sales

Multiply by: your entire target account list. If you have 500 companies, that is 500 keywords and one run.

3. Decision makers by title

LinkedIn public profiles put name, title and company in the page title, and Google keeps that structure.

Pattern: site:linkedin.com/in "[title]" "[industry]" [state]

Multiply by: titles times industries times states. Five titles, four industries, ten states is 200 keywords. You are reading the title column here, not the description. No login, no profile visits.

4. Job vacancies with the salary band

Job boards are protected. Their listing pages are heavily indexed.

Pattern: site:linkedin.com/jobs [role] [city] salary, "[role]" salary posted 2026, site:indeed.com [role] [city] pay

Multiply by: roles times cities. Thirty roles across twenty metros is 600 keywords, and you can rerun the same batch weekly to watch hiring volume move.

5. Real estate prices and specs

Portal snippets are almost always the listing spec line, because that is what the page leads with.

Pattern: [bedrooms] bedroom house for sale [city] price, site:zillow.com condo [neighborhood] price sqft, apartment for rent [zip] 2 bed monthly rent

Multiply by: zip codes. A metro area is 40 to 150 zips. Cross it with property type and bedroom count and one market is a few hundred keywords.

6. Product prices, including resale

Pattern: [model] price, site:stockx.com [model] last sale price, "[model]" price size [n] in stock

Multiply by: your model list. Six hundred sneaker SKUs is 600 keywords and gives you a price sheet across every retailer that ranks, in one run.

7. Software and service pricing

Pattern: [tool] pricing per month, [tool] plans price starting at

Multiply by: your competitor set, run monthly. Small batch, high value, because pricing pages are short and their snippets are unusually clean.

8. Local business phone lists

The Maps scraper is better when you need the full record. When you need a name and a number, search is faster and cheaper.

Pattern: [niche] [city] phone number, [niche] [zip] phone, [niche] [city] contact number

Multiply by: niches times zip codes. This is the pattern that scales the hardest, and the one worth starting with if you want to see the volume for yourself.

9. Company firmographics

Pattern: site:crunchbase.com [industry] companies founded employees, "[company]" founded headquarters employees

Multiply by: your account list or your industry list. Enough to size and rank a target list before you spend anything enriching it.

10. Dates: tenders, deadlines, events, filings

Pattern: [category] application deadline 2026 [state], RFP submission deadline [industry] 2026, site:eventbrite.com [industry] conference 2026 date

Multiply by: categories times states, on a weekly schedule. This one is worth running repeatedly rather than once, because the value is in catching the new posting, not in the archive.

What comes back, and what to do with it

The run gives you four columns per result: position, title, description and URL. The description is where your field lives. The title carries names and roles. The URL tells you the source and lets you dedupe.

Ten thousand rows of that is not something you work by hand, so there are two ways to finish the job.

AI parsing on the description column. Phone numbers, prices, dates and salary ranges are regex-friendly, and if the format varies too much, chain an AI automation and tell it which fields to pull into which columns. One instruction applied across the batch. The one rule that matters: tell it to leave the field empty when the snippet does not contain the value. A dataset that looks complete but is invented in 400 places costs you more than a dataset with gaps.

Send it to Second Brain (brain.hexact.io). Export to CSV and you get a file that ages. Push the results into Second Brain and you get a local database you can query, which changes what the data is for.

Instead of filtering a spreadsheet by the columns you defined at export time, you ask questions after the fact. Which zips came back with the fewest results, meaning underserved markets. Which models moved on price since last month. Which districts still have no email on record. Which roles a company started posting that they were not posting in March.

That last one only works because Second Brain keeps every run with a timestamp. A single batch tells you what exists. The same batch run monthly tells you what changed, and what changed is where the money usually is. Nobody sells against a snapshot.

Where this breaks

I am not going to pretend this replaces everything.

Snippets are capped at roughly 160 characters. You get short fields. A phone number, a price, a date, a salary range, a title. You do not get a full product spec table or a complete staff directory.

Snippets are generated per query, so wording changes results. If the field is not coming back, add the field name to the keyword. phone, email, price, salary, deadline, address. Those words are what Google matches against.

Coverage is partial. You get the districts, listings and products that rank. For exhaustive datasets you still need the source.

The rule I use: if the field is short and the page states it plainly, search it in batch. If you need the full structured record, scrape the source. Most people build the scraper first and never test whether the shortcut would have worked. Test the shortcut first. Twenty keywords takes four minutes.

Run it yourself. First test a sample keyword in Google to make sure it shows what you need. Then create a new workflow in Hexomatic (hexomatic.com), paste the list of keywords into Data input, and add the Google Search Scraper.

Or book the Concierge Service and start with working output instead of setup.

Claude answered. The answer sounded perfectly reasonable. It was also wrong.

Claude did not have the information, so it did what AI does in that situation. It filled in the blanks.

Which is the exact problem we built Second Brain to solve.

Months of telling people to put their important information somewhere AI can query it, and we had never done it for our own product.

That is fixed now. The documentation lives inside the Second Brain knowledge base. It installs with the app and stays updated with it.

Connect Claude to Hexact (Second Brain) desktop app and you can ask:

• What file formats can I import?

• How many files can I import at once?

• What happens if a file is unsupported?

• How does OCR work?

• Where is my database stored?

• How do Stripe subscriptions connect to contacts?

• What does Database Cleanup actually do?

Claude does not guess and does not search the web. It reads the documentation for the version of Second Brain you are actually running.

The better questions are not documentation questions.

Say you have four years of contracts, a Stripe export, a Gmail archive, and several folders of project files waiting to go in. Instead of working out the order yourself, ask:

“What should I import first, and is there anything I should organize before I start?”

Claude answers from how the app actually behaves, including the limits you are about to hit.

The same gap shows up in your business

Say you have three Claude chats open. One on pricing. One drafting an email campaign. One going through customer data.

In the pricing chat, you decide to move a promotion to September.

An hour later, the email chat needs the promotion dates. Does it know what you decided?

No.

So you tell it again, or paste from the other chat, or keep a document updated and remember to attach it every time.

That is the same failure as Claude guessing about tagging. The information exists. There is nowhere for Claude to look it up.

A database is somewhere to look.

Tell the pricing chat to write the decision to Second Brain, and the email chat can query it an hour later. Claude gets three tools through MCP: read the schema, run a query, write rows back. The third one is what makes the first two worth anything across sessions.

Need one customer’s history? It queries that customer’s records. Need a clause out of a 200-page contract? It pulls the sections that mention it instead of loading the file. Need to know how OCR works? Now it queries that too.

Claude does not need to know everything. It needs one place to look.

This gets harder the more data you have

Files in a Claude Project work fine, until there are enough of them that you start deciding what deserves to be in there. At that point you are curating context instead of working.

Second Brain holds years of information and Claude takes only what the current question needs.

No reloading the same files. No paying tokens to reread them. No carrying decisions between conversations by hand.

One database, any number of sessions, the same source of truth for all of them.

Try it

Update or download Second Brain at brain.hexact.io. There is a 7-day trial and no credit card required.

Connect Claude and ask:

“What can I import into Second Brain, and what are the import limits?”

The answer now comes from Second Brain itself.

Then ask it the question you actually care about. The one about your imports, your data, your setup.

If the answer does not cover it, book a video call and ask me instead. You get a real answer, and I find out what is missing from the documentation and add it.

We spent months telling people to put their knowledge into Second Brain. Eventually, we remembered to put ours there too.

Import runs OCR now. Scanned files go in as searchable text with everything else.

How it works

The Knowledge files import card checks your machine and shows a label. OCR ready means scanned PDFs get read during import. Drop them in with the rest. No separate step, no converting first.

The recovered text is treated like any other document. Split into ordered sections, stored as rows, tagged, every word indexed. A 60-page scanned lease behaves like a typed one. Ask about the termination clause and Claude pulls the sections that mention it.

Reading them costs you nothing in tokens

Recognition runs through the OCR built into your operating system. Nothing is uploaded to be read. The scan is processed on your machine and the text goes into your local database.

The alternative is handing each page to an AI as an image. A full page runs well over a thousand tokens before a word of the answer is written, and you pay it again every session, because nothing was stored. Two hundred pages of contracts is a real bill for the privilege of turning a folder into text.

Here the reading happens once, at import. What sits in the database afterward is text, and a query pulls the few sections that matched instead of the pages.

What your machine needs

• macOS 10.15 Catalina or later

• Windows 10 or later. Windows may need a language pack. Settings, then Time & language, then Language & region.

• If your system does not qualify you see OCR unavailable. The rest of the app works as before.

Two limits worth knowing before you point this at an archive. Recognition covers the languages your system supports, and a page in a language your OS cannot read comes back as noise. Printed type on a clean scan lands past 99 percent of characters. Handwriting, faded fax paper and photos shot at an angle can produce errors.

International text stops arriving broken

Detection for non UTF-8 documents is better in this build. Windows-1251, Shift-JIS, the ISO-8859 family and others get identified more reliably at import, so the text goes in readable and comes back searchable.

Also in this release

• Higher import limits.

• Tags apply consistently across a batch and stay searchable next to the content.

• An interface pass, plus better help texts.

Try it on your mixed folder

Update at brain.hexact.io. Free download, 7-day trial, no card.

Find the scanned documents you gave up on, the old client contracts, and drop the folder on the Import page. Confirm it says OCR ready, run the import, then ask Claude about a clause.

Migrating an archive, or running several businesses off one database? Book a call and I will tell you what to import first.

Choose a folder, including everything inside it

There is a Choose folder button now. Pick your top-level folder and Second Brain walks every subfolder underneath and imports what it finds. Nothing to flatten first.

The ceiling for one batch is 100,000 files. I do not expect anyone to hit it. It is there so you stop thinking about the limit. Same for a zip: one zip, nested folders inside, same count. Each file up to 2048 MB. Picking files by hand is still capped at 100, since that path is for when you want three specific documents.

Formats: .docx, .pdf, .txt, .md, .rtf. A zip inside a zip or folder gets skipped.

What happens to them

Each file becomes an entry. The text splits into ordered sections, stored as rows and indexed for search. Tags go on at import, all of them searchable.

So a 40,000-file import is not 40,000 files Claude opens. Ask about a clause in a 2022 contract and it pulls the sections that mention it. How much you imported stops mattering, because you never load all of it.

A one-minute tour

The other thing people got stuck on was the first screen after install. Dashboard, Import, Contacts, Settings, Database maintenance, a connection to Claude, and none of it explains itself.

There is now a tour that walks you through the menus and what each does. About a minute, on the real interface, so by the end you have already touched the pages you will use.

Where to start

Free. Download at brain.hexact.io. 7-day trial, no card. Take the tour, then point the folder picker at your messiest documents folder and ask Claude something you could not answer before.

Annual. Keeping it? The annual plan is discounted against monthly and locks the price for the term. Grab now!

Book a call. Several businesses, a shared knowledge base, or an import that looks complicated? Book a call. I will look at your setup and tell you what to import first.

Everything imports into one local SQL database on your machine. Contacts become rows with columns. Stripe and PayPal history becomes transactions tied to those contacts. Calendar events link to participants by email. Word docs and PDFs go in the same database, not as files sitting in a folder. Nothing is a note about a customer. It is the customer, with keys the software understands.

Claude reaches it through MCP with three tools: read the schema, run a query, write rows back. It sees what tables exist, writes SQL, gets the matching rows. When you want the output kept, it writes into the same tables, so the next session starts where the last one finished.

That means you can ask things nobody built a report for. Which customers paid in 2025 and went quiet in 2026. Who I met with last quarter that never bought anything. What one contact has spent across every payment method I have used.

Documents work the same way

Import splits each file into ordered sections with headings, stores them as rows, tags them, and indexes every word for search. So a 200-page manuscript is not one blob you have to load. It is a few hundred sections Claude can address one by one. It searches for the term, pulls the six sections that contain it, and reads those. Fifty documents behave the same as one, because you are pulling paragraphs instead of files.

What Cowork does

Cowork points Claude at a folder. You describe an outcome and it plans the steps, opens files, runs code, writes results.

Two costs come with that design. Everything it reads passes through the context window, so a question touching thirty files costs thirty files worth of tokens even when the answer lived in one of them. And when it writes, it writes files. It can edit a document or produce a new one, but there is no record underneath to update, so the next session opens everything again instead of picking up where the last one stopped. Second Brain writes rows. Ask it to tag fifty contacts and those tags are in the table, ready for the next query, and no file was rewritten to get there.

Where they split

Selection. A query has a WHERE clause. It returns every matching row and nothing else. An agent reading a folder has to decide which files to open and what to skip once the material outgrows the context window. You never see that decision, and it sets your answer.

Consistency. Same query, same rows, next month too. A pass through a folder depends on phrasing, on what turned up first, on where the window filled. Ask twice, get two answers, both confident.

Tokens. SQL counts, filters, and joins before Claude sees anything. Ask which clients went quiet and it reads 12 rows, not 40,000. The database can be gigabytes while the answer stays a few hundred tokens.

Speed. An index walks straight to matching rows instead of scanning the table. A lookup across 100,000 words lands in milliseconds.

Location. Cowork sessions run on Anthropic’s servers, and local files opened through the desktop app are processed there too. Second Brain keeps the database on your machine. Only the rows a query returned leave it.

I am not going to tell you which one you need. Look at the questions you actually ask in a normal week and judge for yourself which shape they fit. I know my answer for my data. Yours may land somewhere else.

What shipped

Importing files. Better tagging, clearer messages, and faster.

More you can bring in. You can now import emails from a CSV. CSV files from your computer can be up to 32 GB.

Database Cleanup. It now repairs broken links between records and tells you what it fixed.

Faster. Dashboard and Import load noticeably quicker, and Contacts has more ways to filter.

Plus a long list of smaller interface and performance work across settings, contacts, and the import page.

Update at brain.hexact.io.

If you have not imported anything yet, start with one thing. Export your contacts to CSV, or pull the zip from Google Calendar settings, and drop it on the Import page. Then ask Claude something you could not ask before. One query explains this better than I can in writing.

If you want me to walk you through it, book a call. I will show you what to import first for your use case and where the useful queries come from. I run these myself, so bring the hard questions.

Then you ask it a question. “Which of my clients haven’t ordered since January?” Silence. “Which customers are also newsletter subscribers?” Nothing. Your second brain has all the pieces of the answer scattered across pages, and no way to put them together.

That’s not a criticism of Notion or Obsidian. They were never built to answer questions. They were built to store notes. The problem is that “second brain” now describes two completely different things, and most people only know one of them.

What note tools actually are

Notion, Obsidian, Evernote, Logseq. These are filing systems. You write, you link, you tag, you organize. When you need something back, you search for it and read it yourself.

They’re genuinely good at this. If you’re drafting, journaling, building a team wiki, or connecting ideas while you think, a note tool is the right choice. I’m not going to pretend otherwise.

But look at what they store: documents. A note about a customer is a page of prose. The customer’s payment history is a bullet list someone typed three months ago, if anyone typed it at all. There are no records. No fields. No relationships the software understands. A backlink between two notes says “these are related” and nothing more. Related how? Since when? The link doesn’t know.

So when AI plugins arrived for these tools, they inherited the problem. The plugin embeds your notes, finds the chunks that look similar to your question, and summarizes them. Ask “how much did this client pay me last year” and it finds notes that mention the client and money. It doesn’t count anything. It can’t. There’s nothing to count. There’s only text about numbers.

The “just upload it to Claude” path

The other approach people try: skip the note tool, dump your files straight into Claude or ChatGPT. Upload the spreadsheets, upload the documents, ask questions.

For small inputs, this works. A page of text, a few hundred rows. I’ve written about where it breaks in one of my previous articles, so here’s the short version.

Every model has a context window, a hard limit on how much it holds at once. When your files exceed it, the tool doesn’t tell you. It chunks your material, guesses which chunks matter for your question, pulls in the top few, and skips the rest. You never see the selection. When the skipped part was the part that mattered, you get a confident answer built on half your data.

I hit this twice in ways that cost me real time.

First, contacts. Customers in one spreadsheet, leads in another, investors in a third, plus an old CRM export. Different columns, inconsistent formats, duplicates everywhere. Thousands of rows total. I asked AI to merge them. It truncated data, invented entries, and delivered a partial result while acting finished.

Second, my book. 211 pages that needed proofreading. Claude did solid work on the first 20 or 30 pages, then lost track of tone decisions from earlier chapters, hit its output limit, and every “continue” made it worse. Maybe 15% of the book got reviewed, inconsistently.

The AI had the skills for both jobs. The input was in a shape it couldn’t work with. That’s the pattern: uploading files gives AI your data as a pile of text, and a system you can’t see decides which parts to read.

What changes with Second Brain by Hexact

This is the gap we built our product around. It takes a different route: your data lives in a SQL database on your machine. Contacts, payments, subscribers, meetings, notes, scraped pages. Each one a row in a table, each field a column. Claude (or Codex) connects to it through MCP and does something neither a note tool nor a file upload allows: it writes a query and gets back every record that matches.

Ask which customers are two payments behind, and you get every customer who’s two payments behind. Not a summary of the ones the retrieval step decided were worth mentioning. All of them. No hidden ranking, no silent truncation. One approach skims and hopes. The other looks it up.

Both of my failed cases became routine. The contact mess got normalized into one schema, and deduplication became a matching rule on email and phone instead of a guess. The book became 211 rows of clean text, and I could run checks like “list every sentence longer than 35 words” across the whole thing with the same quality on page 211 as on page 1.

The same pattern covers technical documentation, and this one hits engineers and support teams daily. A 400-page equipment manual, an API reference, a compliance handbook, a folder of product spec PDFs. Upload one to a chat and ask "what are the torque specs for this model," and you're trusting the retrieval step to have picked the right chunk out of hundreds of pages. Sometimes it does. Sometimes it answers from the wrong model's section and sounds equally confident. Import the PDFs into Second Brain instead, up to 200+ documents at once through a single ZIP, and every page becomes a searchable record. Now "find every section that mentions this error code," "which procedures require certified personnel," or "compare the safety requirements across these three manuals" run against the full text, not a sample of it. And the answer cites records you can open and verify, instead of a paragraph you have to take on faith.

But exact retrieval is only half of it. The bigger half is joins.

Your data isn’t a pile of separate facts. It’s connected. The person who pays you on Stripe is the same person who reads your Substack newsletter and sat in that Google Calendar meeting last Tuesday. In Notion, those are three unrelated pages, if they exist at all. In a chat upload, they’re three files the model may or may not cross-reference. In a database, they’re linked records, and the questions that cross them are just SQL.

Three examples from my own use:

The newsletter question. For years I couldn’t answer whether people subscribe to my newsletter first and become customers later, or the other way around, and how long the gap is. No single tool has this. Substack knows opens. Stripe knows charges. Once both datasets sat in the same database, I asked the question and had the answer in about five seconds.

The calendar question. Google Calendar knows who I met with. Stripe knows who paid. Neither knows both. With calendar events imported and linked to contacts, I can ask “who did I meet with in the last six months who never became a customer” and get a follow-up list, or “how many meetings did it take before this deal closed” and see the actual count. Your calendar quietly holds your entire sales effort. Alone it’s a schedule. Joined with payments, it’s a conversion report.

The pre-call briefing. Before an important call, I ask Claude “what do I know about this person?” It pulls payment history, subscription status, newsletter engagement, past meetings, and my saved notes into one answer. The CRM alone would have shown me a name and a deal stage. The database shows me that their payment failed twice last month and they unsubscribed from the newsletter. I walk into the call prepared instead of guessing.

No note tool answers these. No file upload answers these reliably. They’re join questions, and joins need a database.

The part nobody else has: data that updates itself

Here’s the limitation shared by every option above, including a static database. The data is only as fresh as your last manual effort. Your Notion notes describe last quarter. Your uploaded CSV was accurate the day you exported it. A lead list is dead the day you download it.

This is where Hexomatic changes the equation. So Hexact (Second Brain) connects to Hexomatic directly, and scraping workflows feed the database automatically while the app runs. No exports, no file management, no pasting into chat.

Set the Google Maps Scraper on a monthly schedule per territory, and every run lands timestamped in your database. “Who’s new in this zip code since last month” becomes a query. So does “who crossed 100 reviews this quarter.” Chain the Page Content Extractor and Emails Scraper off the filtered list, and the unstructured parts, what businesses say about themselves on their own sites, become queryable too.

Same with competitor pricing. A scheduled workflow scrapes competitor pricing pages, the results flow in, and “which competitors raised prices this quarter” gets answered from actual timestamped pages, not from the model’s training data.

One scrape is a snapshot. Months of scheduled scrapes are a dataset that shows movement. That’s the compound effect: your internal records and live market data in one place, and Claude reasoning across both. Which customers in a region haven’t been contacted since a competitor changed pricing there. That question touches your CRM and the outside world at once, and it only works when both live in the same database.

No note tool does this. No chat upload does this. This is the combination.

Honest boundaries

Where the other tools win, they win clearly.

Use Obsidian or Notion when the work is writing and thinking. Drafts, wikis, freeform idea capture. Second Brain is not a writing environment and doesn’t try to be.

Upload files directly to Claude when the job is small and one-off. A single document to summarize, a few hundred rows to clean. Setting up a database for that is overkill.

Use Hexact (Second Brain) when your questions sound like database queries. Counting, filtering, comparing across time, joining sources. The moment your question has a “which,” a “how many,” or a “since when” in it, text storage stops being enough.

The second signal is repetition. Customers, leads, a knowledge base, a market you track. That data grows and changes every week. Uploading it to a chat means starting from zero each time. In a database, nothing resets. Every question runs against everything you accumulated before it.

And it works in both directions. Claude doesn’t just read your database, it writes to it. After a call, say “save this note about the client” and it lands on the contact’s record. Paste a meeting summary, it’s saved. Tell it “this person is now CEO of Acme Corp” and the record updates. No forms, no dropdowns. And writes happen only when you ask. Nothing changes on its own.

You don’t have to abandon your notes to get there. Second Brain imports from Notion, Obsidian, Evernote, Logseq, and OneNote, alongside Stripe, PayPal, Google Contacts, Google Calendar, Substack, email archives, and CSV exports from any CRM. 21 sources total. Your note tool becomes one input among many.

Where to start

Self-serve. Download Hexact Second Brain at brain.hexact.io, 7-day free trial, no card. Import one thing you already paste into AI by hand, a contacts export, your Stripe transactions, your Notion workspace. Connect Claude and ask a join question you’ve never been able to answer. That’s usually the moment the difference stops being theory.

Concierge. Already scraping with Hexomatic, or want to? Send us your brief and our team wires the full pipeline, workflows to database to first queries, through the Concierge Service.

Enterprise. Multiple businesses, multiple territories, a shared knowledge base your whole team queries? Book a call and we’ll architect it end to end.

Here’s what actually happens next. The CSV lands in a spreadsheet. Someone works the top 30 rows. The rest sits there. Three months later the file is stale, half the effort is wasted, and you scrape again from zero.

That’s not a scraping problem. Collection is a commodity now. The Maps Scraper pulls up to 700 listings per keyword, and your competitor can run the same workflow this afternoon. The edge has moved to two things almost nobody does: qualification the spreadsheet can’t do, and timing the spreadsheet can’t see.

The collection pass, in one paragraph

For completeness, because the basics still decide data quality: cover the territory with keyword variations (zip codes, neighborhoods, language variants where the market has them), dedupe by business name if you want unique brands or by address if you want every branch, and cut the list hard before enriching. Review count and website presence are the fastest filters. When we scraped 19,044 local listings for a study, businesses with websites averaged 4.70 stars and 226 reviews against 4.48 and 18 for those without. A website is a proxy for a business developed enough to buy from you, unless no website is exactly the problem you solve. I’ve covered the full collection setup in previous articles, so this piece starts where those ended: with data in hand.

A spreadsheet only answers yesterday’s questions

Here’s the limitation nobody names. A spreadsheet answers the questions you defined at export time. Columns for rating, reviews, website. Sort, filter, done.

But real qualification questions come up while you’re selling, and they don’t map to columns:

• “Which of these have 50+ reviews but no website?” Proven demand, weak digital presence. If you sell websites, marketing, or booking software, that’s not a list, that’s a queue.

• “Who’s busy but slipping? Rank by review count, flag everything under 4 stars.” Visible, high-volume, underperforming. They feel the problem before you call.

• “Which ones mention emergency or same-day service on their sites?” That answer isn’t in any Maps field. It’s in the page text.

• “Which describe themselves as family-owned?” Changes the first line of your email.

• “What service does nobody in this zip code offer?” That’s a market gap, and it’s sitting in the same dataset.

The first two you could brute-force with filters. The rest live in unstructured text: what businesses say about themselves on their own sites. So the enrichment pass matters, and it’s short: chain Page Content Extractor (https://hexomatic.com/automation/get-page-content), Emails Scraper, and Social Links Extractor off your filtered list. The listing told you who exists. The website tells you whether they’re worth your time.

Then send it all into Second Brain instead of a spreadsheet. Connected to Hexomatic, the data flows into your local database as it’s collected, timestamped, tied to the workflow. And now Claude can answer every question above, including the ones you think of mid-campaign, because the database feeds it the right slice instead of you pasting 200 pages into a chat window and watching it silently drop half of them. I’ve written about why that happens; the short version is that models have a hard limit on what they hold at once, and a local database is how you work around it instead of pretending it isn’t there.

Two rules keep the dataset trustworthy. First, Claude structures only what’s in the data: no email on the page means an empty field, not an invented one. A list that looks complete but is wrong in ten places costs you mid-call. Second, write your qualification criteria into a Skill once, and every future batch gets judged by the same standard without re-explaining.

Timing is the part nobody automates

Every sales manager knows the difference between a cold call and a well-timed one. What almost nobody does with local data is systematize the timing.

Local markets move constantly. New businesses open. Ratings slip. A shop crosses 100 reviews. Somebody finally launches a website. Each of these is a buying moment for someone:

• A new business in your territory needs everything: signage, POS, insurance, cleaning, marketing. Whoever calls first wins.

• A rating dropping below 4 means an owner who just started caring about reputation.

• A review count doubling means an operation growing faster than its systems.

None of this is visible in a one-time scrape. It’s visible in the difference between scrapes. Put the collection workflow on a monthly schedule per territory, and Second Brain keeps every run with timestamps. “Who’s new in 33139 since last month” becomes a query. So does “who crossed 100 reviews this quarter” and “who launched a website since our last pass.”

That’s the real upgrade this article is about. A list tells you who exists. A pipeline tells you who to call this week, and why. Same data source, entirely different sales motion.

What this looks like in practice

One monthly Hexomatic workflow per territory: Google Maps Scraper, then Page Content Extractor and Emails Scraper on the survivors of your filters. Second Brain holding the history. A Skill holding your definition of a good lead. And a Monday routine that starts with one question to Claude: “What changed in my territories since last month, and who should I contact first?”

That’s the whole system. No dashboard to build, no CRM migration, no analyst to hire. Which is exactly why it fits a small or medium business: the parts you’d normally pay a team for, the watching, the comparing, the qualifying, run on their own.

Self-serve. Build it in Hexomatic, connect it to Second Brain (brain.hexact.io), and start with one territory and one niche.

Concierge. Or send us the brief: your niche, your geography, your definition of a good lead. Our team builds the workflows and wires the pipeline through the Concierge Service, and you start with working output instead of setup.

Enterprise. Running many territories on a schedule, feeding your CRM and a shared knowledge base your whole team queries? Book a call and we’ll architect it end to end.

It wasn’t being lazy. The truth is more annoying than that. It can’t actually hold everything you give it at once, so behind the scenes it picks. It decides which chunks of your files look relevant to this question, summarizes some of them, ranks what it thinks matters, and leaves the rest out of the answer. All of that happens without you. You don’t get to say “use this part, ignore that one.” The system makes that call for you, every time, and sometimes it calls it wrong.

You’ve felt this even if you never thought about why. You correct the AI on one thing, it fixes that, and in the same breath it drops something it had right a minute ago. You add more context to help, and now it’s juggling so much that it starts missing the parts you actually cared about. It takes one, it misses another. Projects, custom instructions, uploaded files, none of it removes the fact that a system you can’t see is deciding what to pay attention to.

For casual stuff, fine. When the part it skipped is the part that mattered, a customer’s payment status, the one instruction you were clear about, the detail your whole question hinged on, it stops being cute. You’re making decisions on an answer that quietly left things out.

This is the problem Second Brain (brain.hexact.io) was built around, and the fix is boring in the best way.

Second Brain (brain.hexact.io) doesn’t turn your world into a pile of text that the AI then skims and summarizes. It keeps your data as real records in a real database on your machine. Your customers, payments, subscribers, meetings, notes, each as actual rows. When you ask a question, the AI writes a query and pulls the exact records that match it. If you ask which customers are two payments behind, it gets every customer who’s two payments behind. Not a summary of the ones it decided were worth mentioning. All of them. There’s no hidden step ranking your own data and throwing half of it out before you see the answer.

That’s the difference that’s hard to feel until you have it. With a project full of files, the AI is guessing which parts of your stuff to look at. With a database, it’s reading the parts your question actually points to. One skims and hopes. The other looks it up.

And because it’s a real store that you own, two more things fall out of it. It’s complete and it stays. Nothing you put in gets quietly compressed into a sentence and lost. Week one it knows a little. A few months in it’s the most useful thing you’ve got, because everything is still there, still exact, still yours to query. You decide what goes in, you can see all of it, and when the AI answers you know it read the real records, not a sketch it wrote about them.

Here’s when it clicks for people on my calls. It’s the moment they stop trying to write the perfect project description to make the AI behave, and realize the problem was never their wording. The AI was always going to decide what to use. Once the data lives somewhere it has to look things up instead of guess, that whole fight disappears.

Grab Second Brain at brain.hexact.io, 7-day free trial, no card. Import one thing you already have, a contacts export, your Stripe transactions, bank statements, last year’s calendar. Ask it a specific question with a specific answer, like which clients haven’t ordered since January, the kind of thing a project full of files always gets almost right. Watch it return the actual list instead of a confident maybe. That’s the moment the difference stops being theory.

Why this happens, in plain terms. A model can only look at so much text at one time. That limit is called the context window. When your projects, files, and saved memory add up to more than fits, the tool works around it by breaking your material into chunks, guessing which chunks matter for your question, pulling in the top few, and shortening or skipping the rest. Often the guess is good. The catch is that you never see it, so when something relevant gets left out, you have no way to know. A database works differently. Through MCP, the AI runs an exact query against your records and gets back every row that matches, with no guess about what's relevant. It also works both ways. The AI can read from your records and write back to them, so a summary it produces or a note you dictate after a call gets saved into the database instead of vanishing when you close the chat. Your data stays whole and you stay in control of it. That's what Second Brain runs on.

Your competitors publish more about themselves than they realize. Prices, product changes, hiring, positioning, the language they use to sell, the documents they leave on their own servers. It’s all sitting on the public web. The companies that win the game aren’t the ones with secret information. They’re the ones who collect the public information systematically while everyone else checks a few tabs by hand and calls it research.

Web scraping is how you do that at scale. Done well, it turns scattered public data into a structured view you can actually act on. Done badly, it turns into a pile of broken templates you spend more time fixing than using.

This is the full picture: the use cases worth your time, and the two ways to deliver them depending on the job. We run both, and most of this article is about helping you pick the right one before you build anything.

The research use cases that matter most

Scraping earns its place when the manual version is too slow, too shallow, or too easy to skip. These are the jobs where it consistently pays off.

Competitor monitoring. Track pricing, product availability, new features, and messaging across the competitors you care about. The point isn’t a one-time snapshot. It’s noticing the day a rival changes a price, drops a product, or rewrites how they pitch themselves, while it still matters.

Lead generation and prospecting. Build targeted lists of businesses that fit your customer, then enrich them with contacts. For local leads, the Google Maps Scraper is the fastest route: one pass returns business name, address, phone, website, and reviews, so the list is half enriched before you start. For wider, non-local prospecting, the Google Search Scraper pulls relevant companies from across the web, then the Emails Scraper and Phone Number Scraper attach contact details from their sites. Smaller, sharper lists beat giant generic ones every time.

Market and niche research. Map a space you’re entering or expanding into. Who the players are, how they position, where the gaps sit. The Google Search Scraper is built for this wider view: it pulls relevant results from across the web, not just one local area, and the Get Page Content reads the pages it surfaces. Together they give you the lay of the land without a week of clicking.

Person and company research. Due diligence before a deal, partner vetting, sales research before a call. Pull what’s public about a person or a company from across the web so you walk in informed instead of guessing. I’ve written separately about why the first step here is never to ask AI to recall facts, it’s to collect the real data first, then let AI work on it.

Tender, RFP, and bid discovery. Bids and contract opportunities are scattered across portals, agency sites, and procurement pages, each formatted differently. Scraping the right keywords across the right sources turns a daily manual hunt into a feed.

Content and SEO research. See what competitors publish, which topics they cover, and where they don’t. The Page Content Extractor pulls the actual text so you can analyze structure, messaging, and keyword focus across dozens of pages at once.

Public files hiding in plain sight. The Files and Documents Finder uncovers PDFs, decks, brochures, and spec sheets hosted on a company’s own site. Sales material, pricing documents, and product detail that never made it onto a normal page.

Local business data. Beyond single leads, the Google Maps Scraper gives you a full local dataset: names, addresses, phones, websites, and reviews across an area. Useful for market sizing and seeing competitive density in a given location, not just building a contact list.

Every one of these comes down to the same two jobs underneath: collect the data, then structure it into something you can use. The mistake most people make is trying to do both at the moment of collection. That’s what makes scraping feel hard. Separate the two and the right approach becomes obvious.

The two questions that decide your approach

Before you build anything, answer these:

How many sources? A handful of fixed sites, or dozens to hundreds?

How stable and how often? A few predictable sites you’ll re-scrape on a schedule, or a wide, changing list you’ll touch once or a few times?

Few sources, repeated often, stable structure points to one approach. Many sources, broad and mixed points to the other. We’re good at both, and they’re genuinely different jobs.

Approach one: a few stable sources on a schedule. Build the recipe.

You track three or four competitor product pages and a couple of supplier catalogs, and you want fresh prices and stock every morning. The pages are already structured, and you’re going to hit the same URLs again and again.

This is where a custom scraping recipe is exactly right.

When the structure is stable and you’re returning daily or weekly, precise setup pays for itself fast. A recipe in Hexomatic targets the exact fields you want, handles pagination, walks every product, and delivers clean structured rows on a schedule. Price in its column, stock in its column, no cleanup, no babysitting. Build it once and it runs.

Two ways to get there. Build it yourself with the scraping recipe tool: point and click the fields, set the schedule, done. Or hand the brief to our Concierge Service and our team builds the recipe for you, including the awkward sites that fight back. The output is precise and repeatable, which is the whole point when the same numbers feed a decision every single day.

The honest trade: a recipe is tied to that site’s layout, so a redesign means a fix. For a handful of sources that makes sense. Maintaining few recipes (templates) is trivial. It stops being trivial the moment the list grows, which is exactly when the second approach takes over.

Approach two: many mixed sources at scale. Grab the raw page, structure it in Second Brain.

Now you want forty competitors in a niche, a couple hundred local sites, or a long list where every site is built differently and some are product pages too. A custom recipe for each is a non-starter. You’d spend more time engineering and re-fixing templates than using anything they produce.

Here you don’t want a scraper that understands each site. You want one that reads any of them.

Get Page Content, the Page Content Extractor automation, pulls the readable text off any URL. No selectors, no per-site setup. Give it a link, get the text. It works on a pricing page, a service page, a directory listing, an article, because you’re not asking it to understand the layout, you’re asking it to grab the words. Google Search Scraper does the same one level up: a query in, the top results out, ready to feed the extractor. For a large site, the Website Crawler finds the pages and the extractor pulls each one.

One simple step replaces dozens of fragile ones. But raw text isn’t a spreadsheet yet, and that’s the part people get stuck on. The structuring still has to happen. It just happens later, by AI, inside Second Brain.

Where the raw path gets its structure

Connect Get Page Content to Second Brain and the extracted text lands in your local database as scraped pages, tied to the workflow that pulled it. Your data sits on your machine, not on someone’s cloud. Then you ask Claude to turn it into what you need: company name, positioning, pricing signals, key claims, contact details, whatever the job calls for, written into clean columns.

The structure is applied after collection, by AI, to text it can actually read. You get recipe-quality output without writing a recipe for every site. And because text extraction doesn’t care about layout, a redesign on any of those forty sites breaks nothing. The words are still the words. Rerun the same two steps and your sheet updates. This is also how you keep a research dataset current: schedule the collection, and Second Brain stays fresh without you reimporting by hand.

This is the real connection between Hexomatic and Second Brain. Hexomatic collects across many sources. Second Brain holds it locally and structures it at scale. AI sits in the middle doing the one thing it’s reliably good at.

Why not skip Second Brain and paste it all into Claude directly?

Because that’s where AI quietly falls apart, and it’s worth being exact about why. I covered this in one of my previous articles, but the short version: AI is excellent at structuring a small, focused chunk of text and unreliable the moment the input gets big. Every model has a context window, a hard limit on how much it holds at once. Dump fifty pages of raw scraped text into one prompt and it starts dropping records and blending one source into another. The capability is there. The capacity isn’t.

That gap is the entire reason Second Brain exists. The data lives in a local database, and Claude is fed the right slice at the right time instead of everything at once. You get the structuring quality of a small input applied across a large dataset, because the database does the holding and the feeding.

One discipline that protects every dataset

When you ask Claude to structure scraped text, tell it explicitly to use only what’s in the data. If a phone number isn’t on the page, the field stays empty. It does not invent a plausible one. Structure what’s there, leave blanks as blanks. A dataset that looks complete but is quietly wrong in ten places is worse than one with honest gaps, especially when you’re about to make a competitive decision on it.

Mapping use cases to approach

Most price monitoring and product tracking on a fixed set of competitors fits the scheduled recipe: stable sources, repeated often, structured fields you want clean and on time.

Most market mapping, lead research, person and company research, tender discovery, content analysis, and document hunting fits the raw path: many sources, mixed and changing, where the value is breadth and where AI structuring in Second Brain beats building and maintaining a template per site.

Plenty of real projects use both. A scheduled recipe watching your core competitors, and a raw-plus-Second-Brain sweep whenever you research a new market or batch of prospects. The skill isn’t loyalty to one method. It’s reading the job correctly before you start.

Get the right setup for your job

Self-serve. Start in Hexomatic. For a wide research sweep, run Google Search Scraper and Get Page Content across your URLs, pipe the results into Second Brain, and ask Claude to structure them. For a few fixed competitors, build a scraping recipe with the visual scraper and put it on a schedule.

Concierge. Don’t want to build the recipe or wire up the pipeline yourself? Send us the brief through our Concierge Service. We’ll set up either path for your exact use case, the scheduled recipe or the raw-to-Second-Brain flow, and hand you working output.

Enterprise. Running this across many sources on a schedule, feeding internal systems and a structured local knowledge base your team queries with AI? Book a call and we’ll architect the full pipeline around your data, from collection to a research database your whole team can ask questions of.

Your data. The customers, payments, contracts, and notes that no public model was ever trained on. That is the one thing the next founder with the same subscription doesn’t have. Second Brain exists to put it to work.

AI guesses when it should know

Ask a general AI how many customers you have and it makes up a number. Ask what you decided about pricing last quarter and it gives you a confident answer built on nothing. It doesn’t have your records, so it fills the gap with something that reads well and falls apart the moment money depends on it.

Memory features don’t fix this. They store a preference, not a record. There is a real difference between an AI that remembers “I run a SaaS company” and one that can tell you which customers signed up last month, haven’t logged in since, and are still paying. The first is a sticky note. The second is a decision you can act on. And the second only happens when your actual data is sitting somewhere the AI can query directly.

That is the whole idea. AI is excellent at reasoning over your data once it can reach it. It was never a replacement for the data itself.

Your data, your machine, your control

Second Brain is a database that lives on your computer, not in someone else’s cloud. You own the file. You back it up by copying it. Nothing leaves unless you decide it does.

Claude connects to it directly and becomes the interface. You ask a question in plain language and get an answer from your real records. You tell it to log a meeting or update a contact and it writes straight to your database. The key choice: it only touches your data when you’re in the conversation. Nothing runs on its own in the background, nothing acts while you sleep. You stay the operator, not the passenger. For real business data, that is the tradeoff that matters.

Getting data in is the part most tools get wrong, so this one reads the exports you already have. Stripe, PayPal, Notion, Evernote, Asana, Jira, your email archives, folders of contracts brought in hundreds at a time, more than twenty sources in total. You bring the file, it lands as structured, searchable records.

What it feels like once it’s running

The value shows up in the questions you can suddenly answer without building a report.

Instead of “what’s a good pricing strategy,” you get “based on your last three pricing changes and how churn responded each time, here’s what the data suggests.” One is a blog post. The other is built from your numbers and nobody else’s.

The strongest questions cross two layers at once, your own records and outside signals in the same place. Which customers in a region haven’t been contacted since a competitor changed their pricing there. Which clients are due for follow-up based on their last service date. Those only work when internal data and live market data sit together, which is what the built-in scraping keeps current for you while the app runs.

If you sell across channels, the scattered-customer problem disappears. A buyer who ordered on Amazon, then Shopify, then Faire normally lives as three records in three dashboards. Second Brain matches them into one person with full history and lifetime revenue, then answers in plain language: which products sold most last quarter, who hasn’t reordered in 90 days, revenue by channel.

And it shapes itself to you. If you’re not in ecommerce, those tables never appear. The database shows only what holds your data. The product fits your business instead of forcing your business into a template. Your database even travels: it stays on your computer, but you can reach it from your phone mid-meeting while the machine stays home (details here). Only the question and the answer move.

The best part is what happens over time. The data you add today makes the answers better next month, and better still the month after. This is not a tool you use once. It is an edge that compounds every time you feed it.

Where it’s headed

The current version makes data easy to get in and easy to query. Next comes structure: contacts that link to meetings, meetings to decisions, decisions to tasks, so when you ask about a customer the AI connects their payments, the meeting where you discussed them, and the email you sent. More import sources are coming across accounting, email marketing, and ecommerce. An optional assistant layer is in the works, opt-in only, with local-only staying the default because that is the right default and it is not changing.

You don’t need any of that to start. You need one import.

Try it

Take whatever you already paste into AI by hand, get it into the database, connect Claude, and ask a question you couldn’t answer before. That is usually enough to see what changes.

The early-adopter price is locked for life if you join before it ends. After that it goes up and stays up.

Lock in your plan at brain.hexact.io, or book a free walkthrough and we’ll set up your first import with you.

Multiple tags per import

You can now assign multiple tags when importing knowledge base entries, separated by commas. Previously you were limited to one tag per import, which meant either under-tagging your data. Now you can label an entry as “research, competitors, Q3” in one go.

This is useful when a document genuinely belongs to more than one category. A vendor contract might be tagged “legal, vendors, 2024.” A research PDF could be “market research, product, competitors.” Tagging at import time means your data is organized from the moment it lands, not after the fact.

Batch imports: up to 100 files, or 200 in a zip

You can now import up to 100 files in a single batch, or package up to 200 files into a zip and import them all at once.

Here’s where this becomes practical:

• Contracts and legal documents. Drop in a folder of vendor agreements, NDAs, or client contracts and query across all of them. Ask Claude to surface renewal dates, liability clauses, or payment terms without opening each file individually.

• Research and reports. Import a set of industry PDFs, analyst reports, or academic papers and treat them as a searchable knowledge base. Ask questions, get summaries, compare findings across documents.

• Internal documentation. Product specs, SOPs, meeting notes, onboarding materials. Import the whole folder, tag it by department or project, and anyone on your team can query it in plain language.

• Email exports or CRM data. Export a batch of customer conversations or deal notes, bring them into Second Brain, and ask Claude to identify patterns, summarize sentiment, or pull out specific details.

The zip option is particularly useful when you’re migrating an archive or setting up Second Brain for the first time with an existing document library. Package everything, set your tags, import once.

Query your knowledge base from anywhere

Second Brain runs locally on your desktop, but you’re not limited to your desk. By connecting Second Brain to Codex, you can access and query your entire knowledge base from the ChatGPT mobile app. Pull up contract details, search your notes, or ask questions across your documents from your phone, with your data staying local the entire time.

General bug fixes and UI improvements

We’ve also pushed a round of bug fixes and interface refinements based on user feedback. Nothing dramatic, just things that were slightly off and are now right.

If you haven’t tried Second Brain yet, it’s free to download. You connect it to Claude, point it at your data, and start querying it in plain language. No SQL, no dashboards, no setup overhead.

• Download Second Brain: https://brain.hexact.io

• Already using it or have a larger use case or need a custom setup? Book a call with us

Performance and UI improvements

We’ve shipped a round of performance and optimization improvements to Second Brain, along with a refreshed Learn feature dashboard and UI. Queries run faster, data loads more reliably, the overall experience feels tighter, and the Learn section is now easier to navigate and read at a glance.

Unified commerce order history

If you sell across multiple platforms, your customer data has probably always been scattered. Someone buys from you on Amazon, again on Shopify, and they exist as three separate records across three separate dashboards. Second Brain now solves that.

You can import full order history from any sales channel, including Amazon, Shopify, Faire, Walmart, eBay, and more, with support for multiple stores on the same platform. Orders and individual line items land in one place, tied to your contacts. Buyers are automatically matched across platforms by email, phone, or customer ID, so a customer who purchased from you on three different channels shows up as one person with their complete history.

Every platform exports data differently, so we built a universal database structure that can receive any of them. You bring your export file and describe it to Claude or Codex, and it maps your data into the right structure automatically. Once it’s in, you can ask for any report in plain language: which products sold most last quarter, which customers haven’t reordered in 90 days, lifetime revenue by channel. No dashboards to configure, no exports to run.

Contact details in the desktop app now show commerce orders alongside Stripe and PayPal activity, with products, quantities, and prices in one view, and commerce spend included in each contact’s lifetime revenue totals. Stripe and PayPal records now also capture gross, fees, net, refunds, and currency-converted amounts. Multi-currency is supported throughout.

If you’re not in ecommerce, none of this appears in your Second Brain. No commerce tables, no order fields. This is a core principle we’re building around: you only see what’s relevant to how you actually use the product. The Database tab works the same way. It shows only tables that contain your data, and the moment you import something new, that table becomes available to navigate, manage, and search. There’s more to do here, but the direction is set.

Your Second Brain, accessible from anywhere

Second Brain stores your data locally. Your SQLite database lives on your computer, not on a cloud server somewhere, which means you stay in control of it. That’s intentional. But it also raised an obvious question: what if I’m not at my desk?

We’ve added instructions for connecting Second Brain to Codex on the desktop. Once you set that up, you can access your knowledge base from two places: Claude Desktop on your computer, and the ChatGPT app’s Codex tab on your phone. As long as your computer is on, Codex on mobile can reach your local database through that desktop connection.

What that means in practice: you’re in a meeting, you remember you wanted to add a contact or update a note. You pull out your phone, open the Codex tab, and do it. Your computer stays home. Your data stays local. You just gained a remote interface into it.

What you can do from mobile

Anything you can do in Second Brain from the desktop, you can now do from your phone. Query your knowledge base in plain language, add new records, edit existing entries, look something up mid-conversation. The database never leaves your machine. What travels between your phone and your computer is only the query and the response. The data stays where you put it.

How to set it up

The connection instructions are inside your dashboard. Follow the Codex setup guide once, and mobile access works automatically from that point on.

Already using Second Brain? Download the latest version to get all of the above, including the unified commerce history, the refreshed Learn dashboard, and the Codex setup instructions.

New to Second Brain? Start a 7-day free trial and see how it fits into your workflow before committing to anything.

Want a walkthrough first? Book a free demo call and we’ll show you how the whole thing works, including the mobile setup.

I’m learning Spanish. It will be my fifth language. Every time I’ve learned a new one, I’ve started the same way: words. Not grammar rules, not conversation practice, not apps that gamify everything into streaks and badges. Just words, a few minutes a day, as consistently as possible.

Vocabulary is the foundation, and it’s the only part of language learning that’s genuinely measurable. You either know a word or you don’t. You can set a goal like “500 words this year” and actually track it, which is more than you can say for “get better at speaking.”

The problem is I’ve been counting manually. A spreadsheet here, a notebook there, trying different apps, none of them doing exactly what I needed. Most tools are built around flashcard decks or spaced repetition algorithms. Useful, but not the point. I just wanted a simple, honest count: words I know, words I’m working on, progress over time.

Then I started hearing the same thing from Second Brain customers. They were already using AI to learn languages, having lessons with Claude, building vocabulary through conversation. But the tracking was still a mess, manual counts, scattered notes, no real picture of progress. Same problem, same frustration. So we decided to just build it properly.

Introducing Learn

Second Brain now has a Learn section, built specifically for language study. It’s on by default when you open the app.

Here’s what it does:

Language profiles. Add one or more languages you’re studying. Each gets its own word list, phrase list, and progress tracking. You can archive a language you’re not actively working on and bring it back anytime.

Word and phrase tracking. Every word or phrase you add gets a status: new, learning, known, or ignored. When you mark something known, the date is recorded. That’s how your progress becomes measurable over time, not just a feeling.

A real dashboard. You get a progress summary showing total tracked items, known count and percentage, and weekly charts of words marked known over the last 8 weeks. You can filter by the last 7 days, last 30 days, or all time. It’s a simple, honest picture of whether you’re actually making progress.

Claude does the logging. This is the part that changes the workflow entirely. During a lesson with Claude, it can add words you miss, mark ones you already know, log phrases, and update statuses, all through your existing MCP connection. Everything stays on your machine. No cloud sync, same local-first model as the rest of Second Brain.

What This Looks Like in Practice

You open a lesson with Claude and tell it you want to practice Spanish vocabulary. During the conversation, Claude logs new words automatically. After the session you open Learn, check your dashboard, and see exactly where you stand.

You can also ask Claude directly: “How many Spanish words have I marked known this month?” and get a real answer pulled from your data.

A few starter prompts worth trying:

• “Add the top 100 most common Spanish words to my Learn list.”

• “Populate about 50 useful Spanish phrases, greetings, travel, and small talk.”

• “During our lesson, log new words I miss and mark ones I already know.”

• Already have a word list somewhere? Upload your spreadsheet or CSV and ask Claude to insert it into your Learn list. No manual re-entry.

Getting Started

Open Learn from the main menu (it sits after Settings). Go to Profiles and add your first language. Then update your Claude skill from the in-app skill setup page, that’s what gives Claude access to read and write your Learn data. After that, study with Claude or browse and edit your lists directly in the app.

If you don’t need Learn, you can hide it entirely from Settings without losing any data.

Is This for Everyone?

Not everyone is learning a language, and not everyone who does learns through vocabulary lists. That’s fine. If Learn isn’t relevant to you, you can hide it from Settings in one click and nothing changes about how you use Second Brain.

But if you are learning a language and you’ve been tracking progress manually or not tracking it at all, this is worth trying. The AI side was already covered. What was missing was the structure around it.

If you try it and have ideas for what else would be useful, I’d genuinely like to hear them. There’s more to build here.

Download the new version at brain.hexact.io. A 7-day free trial is included, no credit card required.