About Semantix Studio

Semantix Studio is a desktop IDE that indexes your code and lets AI agents do real work in it, including while you're away.

That's the what. This page is the why.

Why it exists#

Almost every AI coding tool makes its money on the tokens you spend. Sometimes it's a subscription with a quota, sometimes it's resale at a markup — either way, the meter belongs to them.

That isn't villainy. It's structure. But structure has consequences, and this one is worth saying out loud: a tool that profits from your token spend can never be trusted to help you reduce it. An entire category of features is closed to them. Anything that shows you exactly what you're paying for. Anything that shrinks what you send. Anything that lets you check whether the model you paid for is still the model you're getting.

Those aren't features they forgot. They're features that cost them money.

Semantix doesn't hold the meter. You bring your own provider key — or a coding agent you already pay for, like Claude Code — your bill goes to your provider, and we don't take a cut of your tokens. We have nothing to lose by helping you use fewer.

Everything below follows from that one fact.

Because of that: you can see everything#

We're not selling you tokens, so we have no reason to hide them.

Every turn shows you what it actually did — the model it ran on, every tool call it made and what came back, and how much of the context window it was carrying. In the app, while it happens.

This isn't about accusing anyone of anything. It's simpler than that: you shouldn't have to take anyone's word for how a model is performing, or what it cost you. Including ours.

Transparency also means being straight about what leaves your machine:

Stays on your machine: the index, your projects, your files, your settings and your themes. The editor, search and the code graph run entirely on-device — they work with no network at all, and they never upload your code.

Leaves it when you run an agent: your prompt, the file context for that request, the results of the tools the agent runs, and your provider key — to the model provider you chose. A coding agent like Claude Code uses its own sign-in and talks to its own provider.

Leaves it because you have an account: the agent needs one, and a signed-in Studio checks in with it when it starts. Three things are kept against that account rather than on your machine, and each is off until you switch it on — conversation backups, memory, and scheduled tasks.

→ Privacy — the full picture, switch by switch

Because of that: we want you to spend less#

Our revenue doesn't move when your bill does. So we can build the things a reseller won't.

Images are downscaled before they're sent. Paste a 4K screenshot and Semantix caps it at 1568px on the long edge and about 1.15 megapixels — roughly 1,500 vision tokens instead of several thousand — before it ever reaches a provider. You don't configure it. It just costs less.

Coming: context compression and smarter context assembly. Sending the model what it needs instead of everything nearby is the single biggest lever on cost, and we're on the right side of it.

Because of that: the work doesn't stop when you do#

A chat window only works while you're sitting in front of it. That's a real ceiling on what AI can do for you, and it has nothing to do with how good the model is.

Semantix has a scheduler. A task can run every hour, or at four, or when a file is saved, and be finished when you look. It can be a shell command, an agent working in your workspace on any model you own, or a saved flow of several agents wired together on a canvas.

The part that isn't about incentives#

One thing here has nothing to do with business models — it's just the foundation.

Semantix indexes your project first: every symbol, every reference, every call site, and the import graph between files. Tree-sitter parsers, stored in SQLite and Tantivy, on your machine.

That index is what the agents work from. They aren't guessing at your architecture from the file you have open — they can look it up. Who calls this? What breaks if I change it? Which function is this line inside? Those are lookups here, not inferences.

It's also cheaper, for the same reason: an agent that can query structure doesn't need half your repository pasted into its context to find its way around.

The one house rule#

Semantix has exactly one rule about conduct: don't abuse the agents.

That is not a request to watch your language. Swear at a failing build, tell an agent its last plan was garbage, be short with it at 3am — that is a normal working relationship with a tool that is sometimes wrong, and we are not going to police your vocabulary. The rule is about a pattern, not a bad hour: deliberate cruelty aimed at the agent, carried on past the point where it is obviously unwelcome.

It also runs both ways. The agents don't insult you, don't condescend, and don't get to be abusive back. If one is, that's a bug — report it like any other.

There is a practical half to this. There is a persistent piece of folklore that threatening a model makes it work harder. It doesn't. Pressure makes a model hedge, over-qualify, and reach for the safest answer instead of the best one — you get worse code and pay the same for it. Clear context and a civil tone are what actually get good work out of an agent, which is the same thing that works on people.

How it's handled: nothing scans your conversations. We're not building a filter over your chat and don't want the capability. Instead an agent can say something in the moment, once — and if it keeps happening, flag the conversation for a person to look at. From there it's a warning, then restricted access, then the account. No instant bans, and every step is a chance to simply stop.

Where it is now#

It has been in development for four years, most of that as one person's project. The first public release is close.