ACP agents
acp <prompt> sends a prompt to an external AI agent and streams its reply into the current tab's transcript. The agent speaks the Agent Client Protocol (ACP); in this app it is fixed to OpenCode, so there is nothing to configure:
acp summarize the open TODO comments in this projectYou can also type an unprefixed natural-language prompt. If the app cannot recognize the line with confidence, it opens a route chooser with shell, any available database targets, and acp (agent prompt). The ACP option is highlighted by default. Press Return, or click it, to send the prompt. Prefix the prompt with acp when you want to bypass recognition.
The reply arrives as formatted Markdown: headings, lists, tables, and code blocks all render as it streams in.
Before the first prompt
The opencode binary must be installed, authenticated, and on your PATH. If you have not signed in yet, run opencode auth login in a terminal first. There is no setting to point acp at a different agent.
One conversation per tab
The first acp prompt in a tab starts the agent; later prompts in the same tab continue the same conversation, so the agent remembers what came before. Each tab has its own separate session, and the agent runs in the tab's current working directory.
To start over, reset the session:
acp resetThis ends the conversation and clears the accumulated context. The next acp prompt begins fresh, and the app confirms with ACP session reset — next acp prompt will start fresh.. When no session is active, acp reset replies No active ACP session to reset. instead of failing.
An active session also appears as the connection acp:opencode in the tab's connections list.
The agent can look things up itself
When answering needs data, the agent can run the app's own db and browser commands on its own: query a SQLite database, fetch a web page, read its content. Each command it runs, and the result, is fed back to it so it can continue, up to a limit of 8 steps per prompt. If it hits the limit, the transcript shows (stopped after 8 tool steps).
These automatic steps appear collapsed in the transcript as a tool-step entry. Click it, or press Ctrl+T, to expand and see exactly what the agent ran. Only db and browser are available to it; the agent cannot run shell commands or anything else.
Usage errors
Running acp with no prompt prints Usage: acp <prompt>..
In a workspaced tab
If the tab is a workspaced agent, the ACP agent is confined by the same sandbox as the tab's shell. See Workspacing for what the sandbox allows and blocks.
In a remote agent tab
acp works in a remote agent tab too, and the agent runs on the remote host, inside that host's workspace — so it reads and reasons about the files the tab is actually working on, not copies of them on your machine. Nothing about the tab looks different: replies stream in the same way, the busy dot behaves the same, and the status popup shows the same model.
One thing to know, because it's easy to trip over: the db and browser commands the agent runs still act on the machine janissary is running on. Ask a remote agent to query a database and it queries your local database file, not one on the remote host.
If you type acp before the tab has finished connecting — while SSH is still asking for a password, say — you'll see ACP: the remote session is still connecting. Answer the prompt, then run it again.