# Connect your AI agent

> For the complete documentation index, see [llms.txt](/llms.txt)

unotest studio doesn't ship an AI model — it **connects the agent you already
use** to the open project, over MCP. The agent drives your real app through the
unotest MCP server and writes plain `.js` tests into the folder.

## Connect a workspace

From **Tools → Connect agents to this workspace…**, studio writes the MCP
configuration — pinned to the current folder — for the agents it supports:

- **Claude Desktop**
- **Claude Code**
- **Codex**

A dialog confirms which were connected. Then open your agent and ask it to record
or run a test: it drives your system browser and writes the result into your
repo.

Claude Code is also wired automatically each time you open a workspace (a
project-scoped config), so running `claude` in the
[docked terminal](/viewer/terminal/) picks up unotest right away.

## You stay in control

Studio is a **provider, not an LLM host** — it never calls a model and never
applies a change on its own, and it holds no API keys. As everywhere in unotest,
the agent proposes, the tests land as readable `.js`, and you review and commit.
See [For your AI agent](/agent/overview/).

## Next

- [Connect your editor](/agent/connect/) — the same wiring, by hand
- [Under the hood](/studio/under-the-hood/)