How to use the AI chat in Studio

When you need this

Use the chat panel when you want an agent to actually work on the project you have open — not just answer questions about it. It stays docked beside every page of the project, in both Flow and Research mode, and it runs your own local install of Claude Code or Codex: never a hosted API, and Studio never reads or stores your credentials itself. Use it to ask for an edit, hand it a bug to chase down, or drive the whole project the way you’d drive it from a terminal, while watching every command, file edit, and tool call as it happens.

Steps

• Pick a backend. Studio offers exactly two: Claude Code and Codex. You choose one during setup and can change it later from Studio’s settings; whichever is selected, Studio launches your own local install of it and signs in with whatever session that install already has — it doesn’t ask you for an API key.
• Start or switch threads from the dropdown at the top of the panel. New thread starts one against the flow you currently have open; the dropdown also lists every thread you’ve already started, labeled with its backend and flow, so you can jump back into an earlier conversation instead of losing it when you open something else.
• The model, effort, and approval-mode menu is one popover, opened from the button next to the composer. It lists the models your backend’s catalog offers, a reasoning-effort slider for models that support more than one level (Low through Max), and four approval modes:
  • Manual — the agent asks before it changes anything.
  • Edit automatically — the agent edits files without asking.
  • Plan — the agent explores and proposes, without editing.
  • Trust — the agent runs everything without asking, including commands.
• Changing this menu on an open thread changes that thread immediately — it’s not a setting saved for next time. Switch the model, effort, or approval mode while a thread is open, and Studio applies it to that thread right away, mid-conversation. Only when no thread is open yet does your pick become the starting point for the next one you create. The menu’s button always shows what the open thread is actually set to, not a choice you made but haven’t applied.
• What the agent can do inside your project: run shell commands, showing the command and its exit code; edit files, showing a diff of the change; and call tools, including tools from any MCP server connected to the project — shown as the server name and tool name together, with its arguments and result.
• An approval gate shows up as an Allow/Deny bar right on the tool call it belongs to. Whenever the current approval mode requires your decision before a command runs or a file gets written, the tool’s card grows a bar with Allow and Deny buttons and a short line saying what it’s asking permission for. Nothing happens until you answer.
• Keep writing while the agent works. The text box and Send stay available during a turn: press Enter or click Send, and the message shows up at the bottom of the thread, marked Queued — the agent reads it at its next step. The agent picks it up as soon as the command or tool call it’s running finishes, and from that point the message sits in the conversation like any other you wrote. If the agent was already writing its final answer, it reads your message right after that answer ends, as a new turn. A message marked Queued — sent after this turn waits for the current turn to end and then starts the next one: Codex does this when it can’t take new input yet, for example in the first moment of a turn.
• Long threads open fast and look exactly as they did live. Opening a thread, switching back to it, or reloading the page shows its newest turns right away: every thinking block, every tool call with its full arguments and result, commands with their output, file edits with their diffs, and the answers — nothing is summarised or left out. Scroll up and earlier turns load by themselves as you get near the top; there’s no button to press. New activity keeps streaming in at the bottom as usual.
• Stop a running turn with the square button next to Send. It stops the agent where it is. Messages you queued aren’t dropped — the agent reads them next, in a new turn.
• If Stop doesn’t take, use Reset. When the thread still shows Working… a few seconds after you pressed Stop, the square button turns into Reset. Reset doesn’t wait for the agent: Studio closes the agent’s process, marks the turn stopped, and unlocks the composer. Your next message starts the agent again and it carries on with the same conversation.
• The agent can carry on by itself. If it started a command in the background during a turn, it may wake up after that turn ended — when the command finishes — and keep working. That work shows up as a new turn you didn’t write a message for, opened by a line reading The agent continued on its own, and Stop works on it like on any other turn.
• A server stop or restart never leaves a thread stuck. If the Studio server stops while a turn is running, the turn shows as stopped as soon as the server is back, instead of Working… forever. Send a message to go on; the agent resumes the conversation where it left off.
• The agent can’t stop or restart the server it runs in. A command that would kill the project server, or start another aqven dev or aqven serve, is refused and the agent is told why — that server is what keeps the chat alive. It never needs a restart to use new code: the next run picks up whatever the agent or you changed in the project’s Python files — code steps, tools, generated types.
• The Add context button next to the composer and each message’s more actions (⋯) button don’t do anything yet — they’re placeholders reserved for later.

Example

Open the showcase project’s support_case flow and its chat panel. Start a new thread, leave approval mode on Manual, and ask: “Add a not_empty check on the reply text to the reply_look experiment.” The agent reads experiments/reply_look/experiment.yaml, then proposes writing the change. Because the thread is in Manual mode, an Allow/Deny bar appears on that edit before anything is written. Click Allow, and the diff lands in the file.

Now open the model menu and switch approval mode to Trust without starting a new thread. Ask it to run aqven check to confirm the experiment file is still valid. This time no approval bar appears — the mode change applied to the thread you already had open, so the same thread that just asked permission now runs the command on its own and shows you the exit code.

See also

• How to set up a coding agent outside Studio — the same setup for Claude Code, Codex, or another agent you start yourself instead of from this panel.
• How to use Studio settings — where the chat backend is chosen and switched.
• How to check a project before committing — the command the example asks the agent to run.
• How to use Research in Studio — where a check you asked the agent to add shows up once a series scores it, and where Suggest hypotheses hands the chat its next task.
• How to read a workflow’s graph — the canvas the chat panel sits next to.