Semantix is an AI-first IDE where agents do the work. Schedule it, automate it, compose it from parts, and every agent keeps a memory of its own. Conversations become commits.
Your models
Paste one API key and a provider's whole model list shows up in the picker — any OpenAI-compatible or Anthropic endpoint, local or hosted. Already paying for Claude or ChatGPT? Your subscription signs in once and the agents run on it. Living in opencode, Copilot or another CLI? Semantix drives it as the agent. Semantix sells the workspace, not the tokens.
Watch your agent work
Nothing staged in these shots: a real session in a real project, captured as it ran. We build Semantix in Semantix the same way.
Ask for a change and the agent's first moves light up the map of your code: every folder it lists, every file it opens, glows on the graph, so you see what it read before it touches anything. Then each file it writes opens as a live diff in the chat while a card on its node types the code in, line by line, with the token count ticking. Nothing hides behind a progress bar.
Many agents
One fixing a bug, one writing a feature, one reviewing the last commit, one onboarding to a new repo. All at once, side by side, on the same map of your code.
Every chat floats as its own window over your code, with its own conversation, its own model and its own work in flight.
Minimise one, maximise another, leave the rest working in the corner of your eye. Each window shows what its agent is doing right now.
There is no agent count to buy. Open as many as your hardware and your model plans can carry.
Task Manager
A chat window only works while you sit in front of it. A task doesn't. Write what you want done the way you'd tell a colleague, pick the project and the agent, and set the clock — every morning, every push, or just once at three.
Every run leaves a record you can open: what the agent read, what it changed, what it decided not to touch, and why. Nothing lands in your tree without a trail. If a run got stuck, you see exactly where — and you pick it up from there.
Agents
The Agent Composer is a canvas. Drop the pieces, wire them together, give them capabilities, press Run. What you draw is what runs.
Drop an agent on the canvas. Wire in a trigger, an output, a second agent in parallel or in a loop. The drawing is the program.
Reading and writing files, searching your code, running the type-checker, running a command. Two clicks each, and the agent has hands.
Every call streams onto the canvas as it happens. Save the flow, run it again tomorrow, or hand it to a task and let it run on its own.
Memory
Close the chat, open a new one tomorrow, and it starts where you left off — your stack, your conventions, the reasons behind them. No notes to keep, no context to paste back in, no “as I mentioned before.” You just carry on.
Memory belongs to a named agent of yours and goes wherever you let it work. Open the Memory tab and see that mind as a sky of lights — orbit it, search it, open one memory at a time.
Free isn't a trial. The whole Studio is $0: the 3D graph, the agent crew, 10 tasks with 15 hours of runs a month, and 1,000 memories across five agents — about three months of heavy use, and nothing you write is ever lost. Your own keys, no card.
Then press ⌘K and say hello.
