Jackson Ly

Building recal in public: a local-first personal AI for macOS that learns how you work and starts doing it for you. Early access below ⬇️

[Jackson Ly @jacksnxly]

I built an AI that watches everything I do.

And 0 bytes ever leave my Mac.

Every tool I tried promised to do my work. Each one just became more work. Another thing to set up, prompt, and babysit. None of them actually knew how I work, so the same tasks kept running through me. Again and again.

So I built the opposite.

recal quietly watches how you actually work, on your Mac, on-device. It learns your patterns, starts doing the busywork itself, and surfaces the finished result for one tap. You approve it, or you don't. Nothing ships without you.

And it's private by architecture, not by policy: the model runs on your Mac. Your activity, the index, the work it does, all stay on-device. No account. No server. Pull the ethernet cable and it still runs.

The end of doing. The beginning of directing.

recal. Your work, autonomous.

Join the waitlist → https://t.co/7yugUBsrR3

[Jackson Ly @jacksnxly]

Most builders would ship better AI tools if they studied how people actually work more than how people describe how they work.

[Jackson Ly @jacksnxly]

Most AI agents are built on a flawed assumption: that your expertise lives in what you can describe.

It doesn't. It lives in what you actually do, the micro-decisions you make without thinking, the shortcuts you've built over years.

So we keep trying to prompt our way to automation. Write better instructions.

Be more specific. And it still misses the point.

The real signal is behavioral. Watch someone work for a week and you learn more about how they think than any prompt ever captures.

That's the bet with Recal. Not a chatbot.

Not a prompt interface. Just passive observation, pattern recognition, and drafts you approve or kill.

Your workflows already contain the intelligence. We just haven't had a way to read them.

[Jackson Ly @jacksnxly]

The most valuable skill for indie builders right now is knowing how to automate your own work.

If you want to get sharper at it, here are the resources that actually changed how I think:

1/ The Personal MBA - Josh K.
2/ Getting Things Done - David A.
3/ Deep Work - Cal N.
4/ The Lean Startup - Eric R.
5/ Shape Up - Ryan S.
6/ Atomic Habits - James C.
7/ The Mom Test - Rob F.
8/ Zero to One - Peter T.

Read these. Then go build something.

What would you add?

[Jackson Ly @jacksnxly]

I don't fully agree with every tool I use. But I've seen builders dismiss an entire approach just because one implementation was clunky.

Perfect alignment is a strange thing to demand from software, let alone people.

[Jackson Ly @jacksnxly]

Do not build what others expect by chasing problems that are not yours. Remember that the thing you are quietly obsessed with solving is exactly the thing worth building.

[Jackson Ly @jacksnxly]

If you are choosing between building what users ask for and building what you see them actually do, build what you see them do.

Because if asking them were enough, you would have shipped the right thing already.

[Jackson Ly @jacksnxly]

NEVER TELL PEOPLE THESE THINGS
1- What you're actually building next.
2- Which tools you've already ruled out.
3- The assumption you can't shake.
4- How close you are to quitting.
5- What your real unfair advantage is.
6- Who you're watching closely.
7- The version you killed before launch.

[Jackson Ly @jacksnxly]

@signulll they also compound. a weak story makes people read every rough edge as incompetence; a tight one buys patience for the gaps you haven't closed yet. you can't polish your way out of a product nobody can explain in one line.

[Jackson Ly @jacksnxly]

Love the ambition. One thing that bites everyone chasing "remember everything": total recall quietly swaps one problem (I forgot) for a worse one (everything comes back, unranked). The leap that actually matters is the system knowing WHEN to resurface something and letting you approve it. Memory you can't steer is just a second feed to drown in.

[Jackson Ly @jacksnxly]

Agree on the structure, with one caveat: typed links and metadata are you predicting what future-you will need, and that prediction rots. The signal that ages best is behavioral, which notes you actually open together, around what work. Hand-authored structure plus observed usage beats either alone, and the second half is the part almost nobody captures.

[Jackson Ly @jacksnxly]

Love this framing. The jump nobody talks about is 4 to 5: once it "runs while you sleep" and surfaces connections on its own, capability stops being the bottleneck and trust becomes it. A system that resurfaces things unprompted only works if you can see why it chose that one and veto it. Retrieval you can't audit is just a fancier kind of noise.

[Jackson Ly @jacksnxly]

This is the part wrapper pricing never wanted you to notice. When the model runs on your machine, the model stops being the moat. All that's left to sell is whether you actually made the workflow better, and most wrappers didn't. Local-first doesn't just save the fee, it removes the excuse. (building recal on exactly this bet)

[Jackson Ly @jacksnxly]

The export-control version is the loud one. The quiet version hits everyone: price hikes, a deprecated model that breaks your prompts overnight, a rate limit, a ToS change. Same lesson. If the capability lives on your machine, none of that is someone else's call. Building recal fully local for this reason.

[Jackson Ly @jacksnxly]

Most teams building agents don't fail because they can't capture knowledge. They fail because they capture it, then immediately try to prompt their way to something smarter.

Rather than going deeper on what they already observed, until it actually works.

[Jackson Ly @jacksnxly]

Observe a workflow.
Find the repetitive part.
Build something that removes it.
Show people what you removed.
Let them remove theirs too.

Do it for yourself and it's called productivity.
Do it for others and it's called a product.

[Jackson Ly @jacksnxly]

the biggest "workflow hack" is to fix how you actually work before you try to automate it.

you will never get much done by prompting an AI if you don't understand your own patterns first. the tool can't see what you haven't noticed yet.

[Jackson Ly @jacksnxly]

Observe a workflow.
Find the repetitive part.
Build something that removes it.
Show people what you removed.
Let them remove theirs too.

Do it for yourself and it's called productivity.
Do it for others and it's called a product.

[Jackson Ly @jacksnxly]

agreed, and the line you drew is the right one. the useful version isnt better storage or a prettier graph, its the system actually DOING something with what you captured. the second the AI is just reorganising the pile, youve built a fancier filing cabinet. the value shows up when it acts on the notes and you just approve.

[Jackson Ly @jacksnxly]

the thing i'd optimise for isnt which app, its never being stuck here again. a decade of notes hostage to one vendors roadmap is the real problem. keep them as plain markdown files on disk and the app becomes swappable. bonus: thats also what claude code/cowork works with best, it just reads and edits the files directly, no MCP needed. obsidian sits right on top of a plain-files vault.

[Jackson Ly @jacksnxly]

the feature isnt the scary part, the destination is. "always-on capture" that processes and stays on your device is just a memory you own. the same capture streamed to googles cloud is surveillance. identical capability, opposite trust model. the real question is never "does it listen" but "where does the audio actually go."