Chat

The panel on the right is where you work with the agent. It looks like a chat, and the first thing to understand is that it isn't one.

A chatbot answers from what you paste into it. This agent works inside your project. It reads your files, searches the index, runs your type-checker, writes and edits code, and runs commands in a real terminal — on your machine, in the project you have open. The conversation is the interface; the work is the point.

Almost everything else on these pages follows from that one difference.

Because of that: who is allowed to change your files matters#

An assistant that can delete files is a different proposition from one that can only talk. Semantix has a permission gate for exactly that: reading and searching happen freely, and everything that writes, deletes, moves or runs a shell command stops and asks you first, naming the tool and showing the target. Allow once, allow that class of action for the session, or refuse.

That gate applies to conversations running on your own OpenAI-compatible model, where Semantix executes the tools itself. On a Claude Code conversation — which is the default — the Claude Agent SDK runs its own tools, and Semantix renders what it did rather than executing it. The gate is not in that path.

The automatic workspace checkpoint — the thing the Restore control on a message rolls back to — is taken by that same gate, so it exists on custom-model conversations and not on Claude Code ones.

→ What the agent can do

Because of that: you can watch it work#

If the agent is editing real files, you should be able to see it happen rather than read a summary afterwards.

So the file it's writing can open in a tab and type itself out; the 3D view of your project lights up the nodes it touches and flies the camera to them; and a line above the message box narrates the current step. Every tool it calls leaves a card in the transcript with a clickable filename, so checking its work is one click, not a search.

→ Watching it work

Because of that: one conversation is often not enough#

Real work runs in parallel — one agent on a migration, another on a bug, a third answering a question. Chat mode shows you one conversation at a time in the side panel. Runs mode slides that panel away and turns every open conversation into a floating card over your project, all live at once, driven by a single command bar.

Same conversations, same history. A different way of looking.

→ Runs mode

The part that isn't about the agent#

Your conversations are stored on your own disk, in a database that lives outside the project folder so a reindex can't take it with it. Semantix does not keep a copy of your chat history on a server unless you switch cloud backup on yourself — it's off by default.

The turns themselves do go out, because that's what running a model means. Where they go depends on which model you picked: a Claude Code conversation goes through your own local Claude login, and a custom OpenAI-compatible model goes to your endpoint with your key.

→ Conversations and history

Getting started#

You need to be signed in, and you need a project open — tools don't run without one.

Type in the box at the bottom, press Enter. Drag a file onto the box from the explorer to point at something specific, paste an image to show it something, type / to reach your saved commands.

• Your first conversation — sending, reading a turn, and finding your way around a long one.
• The composer — attachments, slash commands, choosing the model, and the context ring.
• What the agent can do — the tool list and the permission gate.
• Watching it work — the indicator, follow modes, and what stopping actually stops.
• Runs mode — floating cards, arranging them, and what persists.
• Conversations and history — tabs, replay, storage, custom models and instructions.