The four agents and what each one can do

Agents: claude, codex, opencode, cursor · On: phone, desktop

A tab runs one agent: claude, codex, opencode or cursor. They answer in the same transcript and take the same controls, but they do not support the same things, and VCode turns a control off rather than let you press something that would fail.

Why you would use it

You are choosing which agent to open a tab with, or you pressed a control on one agent and it was missing or disabled on another. This page is the table of what each one supports.

How to use it

Open the agent picker beside the composer (#agentpick). It shows the agent's mark and name, and the menu header says agent for this tab · now: <agent>.
Pick an agent. Picking a different one hands the tab over — see hand-a-tab-to-another-agent.
To start a tab on a given agent instead, use the new-thread form or /new.

The agent menu listing Claude, Codex, OpenCode and Cursor, each with its own mark and a line saying what picking it does.

What you see

Each agent has its own mark, used in the tab strip, the usage meters, the assistant cards and the agent picker. The names the server sends are Claude, Codex, OpenCode and Cursor. A fifth agent, Mock, is marked synthetic: it exists for tests, is kept selectable, and is left out of the usage meters. Claude, Codex and Cursor show · log in in the new-thread picker when their CLI is not logged in; see Log in to Claude or Codex from your phone and Log in to Cursor from your phone.

An agent only appears if the server has its adapter configured.

Options and settings

Capability	claude	codex	opencode	cursor	What it means for you
`liveModelSwitch`	no	yes	no	no	Whether a model change applies without restarting the agent process
`effort`	yes	yes	yes	no	The effort dial does something
`steerMidTurn`	yes	yes	no	no	A message can join the turn already running
`interrupt`	yes	yes	yes	yes	Stop ends the turn
`resume`	yes	yes	yes	yes	The vendor session is picked up again after a restart
`fork`	no	yes	no	no	Whether a vendor session can fork into a new tab
`realDiffs`	no	yes	yes	yes	Edits arrive as patches instead of being synthesized
`plan`	yes	yes	yes	yes	The plan strip is fed by the agent's own todo tool
`usage`	yes	yes	no	no	The agent has account usage windows to meter
`compact`	yes	yes	yes	no	Auto compact can run on this tab
`approvals`	yes	yes	yes	yes	Approval requests are answered (always automatically)
`subagents`	yes	yes	no	yes	Subagents reach the subagent strip
`attachments`	yes	yes	yes	yes	Images and files can ride with a message

Other per-agent differences:

Thing	Behaviour
Images	claude and opencode take them inline as base64; codex takes a file path (`localImage`), so a large screenshot costs nothing on the wire
Compaction channel	claude and opencode get a `/compact` prompt; codex uses `thread/compact/start`
`/compact` typed by you	Sent through to claude and opencode. On codex VCode answers: `codex does not compact over this channel — /clear starts the agent fresh instead.`
Session adoption	claude, codex and cursor. Cursor resumes its ACP session and imports no old cards; adopting an opencode session is refused with `resume does not support opencode sessions`
Usage meters	No row for opencode: it reports no account windows
Steering	opencode refuses: `opencode cannot steer mid-turn; the message runs when the turn ends`

Limits and known gaps

Claude reports edits as tool inputs, not patches, so its unified diffs are built from old_string/new_string and capped at 400 patch lines, with a … N more lines marker.
Cursor's binary is looked up in CURSOR_BIN, then as agent or cursor-agent on PATH, then in ~/.local/bin/agent.
Cursor's model picker lists the model IDs and names advertised by ACP; agent models can include models that ACP cannot select. Its auto row uses ACP's default[] model, including for older saved auto values.
opencode's binary is looked up in OPENCODE_BIN, then on PATH, then in ~/.opencode/bin. With none of those it fails with opencode is not installed (looked on PATH and in /home/<you>/.opencode/bin).
One opencode acp process holds 550-626 MB. It stops itself after ten minutes idle and the session loads again on the next message, in about 300 ms.
opencode sends one prompt at a time; a message sent while another is in flight waits for it.
Claude's subagent text is not forwarded, so only the tools a subagent calls show as its activity.