Working with the agent
Attach what you're looking at, use suggestions, follow what the agent does in Studio, approve its requests, and manage chats.
Once an agent is set up, open the panel with Agent in the top
bar or ⇧⌘J (Ctrl+Shift+J on Windows and Linux). Closing the panel doesn't stop the chat. The
Agent button keeps a dot while a chat is running, and the dot turns amber when the agent is
waiting for an answer from you.
Starting a chat
The agent starts when you send your first message. A line at the top of the chat says Studio's tools are connected; open it to see the list. If Studio's MCP server wasn't running, the chat started it for you.
An empty chat suggests up to four things to ask, based on what you have open. For example:
- send the operation on screen and check the response against the spec
- test the open API against a local mock
- build a collection that walks through the API's main flow
- explain why the last run of the open collection failed
Click one to send it, or type your own.
The Studio context chip
Your first message carries a short note telling the agent that it's working inside Studio, which tools it has, and what's open: the API and its version, the operation, and the collection with how its last run went. The message shows a Studio context chip. Open it to read exactly what was sent. Secret values are replaced, as everywhere else.
Attaching what you're looking at
Type @, or press the @ button, to attach:
- the operation on screen, or another one from the open API
- one of your last five responses
- a step from a collection
- one of the last three run logs of a collection
Use the arrow keys to move through the list and Enter or Tab to pick. Attachments are sent as
text, with secret values replaced by their {{name}}, and show as chips on your message.
Following what the agent does
Each tool the agent uses shows up in the chat as a card, with what it asked for and what came back.
- When the agent creates a collection, it appears in Collections as a draft while the agent builds it, with a banner that names the agent. Keep makes it yours; Discard deletes it, after you confirm. Agents can change only their own drafts. See Collections agents build.
- When a collection runs, the status bar shows its progress, such as "Running Orders lifecycle · step 2 of 5". Click it to open the run. When the run ends, a notice says whether it passed, or which step failed and why, with View run. A passed notice goes away after a few seconds; a failed one stays until you close it.
- The card for a collection the agent created, or a run it started, has Open in Studio, which takes you to the collection or that run's log.
A run the agent starts doesn't take you away from the chat. Only runs you start from a collection's own page open their log when they finish.
Approving what the agent does
There are two kinds of question, and they look different on purpose.
- Studio's approvals appear in a Studio dialog. Studio asks before the agent's first request to
each host, before a
DELETE,PUTorPATCHto a host that isn't on your computer, before it starts a local mock, and every time it saves a collection to a file. You can answer Don't allow, Allow once or Allow for this session. An approval nobody answers is refused after five minutes. See Approving what agents do. - The agent's own permission prompts appear in the chat, with the choices the agent offers. They come up when the agent wants to use one of its own tools, for example to read a file in its folder. Studio doesn't give the agent access to your files or a terminal itself.
Stop, end and start again
- Stop stops what the agent is doing now. The chat stays open, and you can send another message.
- End chat stops the agent. The chat is kept, and you can read it in Past chats.
- New chat ends the current chat and starts a fresh one.
Quitting Studio stops every agent it started.
If Studio's MCP server stops or restarts while a chat is open, for example after Stop agents in the MCP tab, the chat says the agent no longer has Studio's tools and offers Reconnect. If the server couldn't start, it offers Start it and reconnect. Reconnecting starts a new session with the agent, which doesn't see the earlier messages.
Past chats
Studio keeps your last 30 chats on this computer, with secret values replaced. Open Past chats at the top of the panel to read one again or delete it.
A past chat can be read but not continued. Sending a message starts a new session, and the agent doesn't see the earlier messages.
Resizing the panel
Drag the panel's left edge to make it wider or narrower, or focus the edge and use the arrow keys:
Home makes it as wide as it goes and End as narrow. Studio remembers the width. The page next to
it always keeps enough room for its controls.
Keys in the chat
| Keys | Action |
|---|---|
⇧⌘J | Open or close the panel (Ctrl+Shift+J on Windows and Linux) |
Enter | Send the message |
Shift+Enter | New line |
@ | Attach an operation, a response, a step or a run log |
↑ ↓, then Enter or Tab | Choose an attachment |
Esc | Close the attachment list |
Set up an agent
Choose Codex, Claude Agent, Gemini CLI or another ACP agent for the chat panel, sign in to it, and choose the folder it works in.
Studio's instructions
What Studio tells the agent about working in Studio, the task guides that come with it, where they're written, and how to change or turn them off.