Agent chat

The Agent panel is a chat with a coding agent that's on your computer. Studio doesn't run a model of its own. It starts the agent you choose, talks to it over the Agent Client Protocol (ACP), which is how editors such as Zed host agents, and shows what comes back.

Press Agent at the right of the top bar, or ⇧⌘J (Ctrl+Shift+J on Windows and Linux). The panel opens on the right, and the page next to it gets narrower, so nothing is hidden behind it.

What the agent can do in StudioCopy Anchor Link

When a chat starts, Studio connects the agent to its local MCP server, so the agent can use your APIs the way you do:

• look up the APIs in your library, their operations and their specs
• send requests, with your environments and auth, and check each response against the spec
• start a local mock of an API and send requests to it
• build collections, run them and read their run logs
• check a running server against its spec

Studio asks you before the agent's first request to each host, before it starts a mock, and before anything else that has an effect outside the chat. The agent never sees secret values: it writes {{token}} and Studio fills in the value when it sends the request.

The agent also knows it's working inside Studio. The first message of each chat tells it what you have open, and Studio gives it a short set of instructions on how to work here: start from the spec, use a local mock for anything that changes data, and build tests as collections you can keep.

What it looks like in useCopy Anchor Link

Say you have the Orders API open and you ask the agent to test it end to end.

1. The agent reads the Orders API's operations through Studio's tools.
2. It asks to start a local mock of the Orders API. Studio shows you the request, and you allow it.
3. It creates a collection, which appears in Collections as a draft while the agent adds steps: create an order, read it back, cancel it, then expect a 404 when it reads the cancelled order.
4. It runs the collection. The status bar shows which step is running, and a notice tells you whether it passed.
5. In the chat, the agent sums up what passed and what didn't, with the field and what the spec says for each mismatch. The tool card for the run has Open in Studio, which takes you to the run log.

If you like the collection, press Keep on its banner and it's yours. Agents can run a collection you've kept, but they can't change it.

What goes whereCopy Anchor Link

What you type and attach goes to the agent you chose, and from there to its model provider. Secret values are replaced with their {{name}} first. Nothing else leaves your computer because of the chat, and Studio keeps your past chats on this computer only. See Privacy and safety.

In this sectionCopy Anchor Link

• Set up an agent: Codex, Claude Agent, Gemini CLI or another agent, and the folder it works in.
• Working with the agent: attachments, suggestions, approvals, runs and past chats.
• Studio's instructions: exactly what the agent is told, and how to change it.
• Privacy and safety: what leaves your computer, and what the agent can and can't do.
• Troubleshooting: when an agent isn't found, isn't signed in, or doesn't use Studio's tools.
• Contributing: suggest better instructions, or report something an agent did wrong.