# Agents

Which agents brainplane works with, how it reaches them and how work moves between them.

Canonical: https://brainplane.app/docs/agents/

## Supported agents

brainplane works with 14 agents: Claude Code, Codex, Grok, OpenCode, Pi, Omp, Amp, Cursor CLI, Gemini CLI, Antigravity CLI, Copilot CLI, Qwen Code, Factory Droid and Aider. Claude Code, Codex and Grok are tested live on real accounts.

Any tool that speaks MCP can read and update the shared brain through the brainplane MCP server.

## How brainplane reaches them

brainplane runs the official, unchanged command line tools on your own logins: `claude -p`, `codex exec` and `grok -p`. It passes the context pack on standard input and reads the result. Your agents keep their own updates, settings and logins.

It works with agents where they already run: iTerm2, Terminal, tmux and herdr. No session has to move.

## The MCP tools

Each agent gets 10 tools: find the right topic, read its brief, update its state, log a decision, ask a question, raise a stop, register a draft, post a card, report a handoff as done, and link its session to a topic. Agents can open cards. Only you can answer them.

## Handoffs

A handoff moves work from 1 agent to another under a contract: a goal, a context pack, a time budget and acceptance checks. The checks run before the result reaches you. A failed check sends a fix list back to the agent that did the work, with a capped number of rounds.

| Mode | What the agent may do |
|---|---|
| review | Read only. Codex runs in its read-only sandbox, Claude Code and Grok with a read and search tool list |
| research | Read and search the web |
| write | Edit files, within the agent's own permissions |
| browser | Drive the browser, 1 agent at a time, with a queue |

The result lands as 1 finished card with the verdict, the checks, where the result is and the command to resume the session.

## A second opinion

Hand a result to another model family for review, for example a Claude draft to Codex. Both run on your own plans. The review card shows the verdict and the check list.

## Answer a running session

1 key answers a live Claude Code, Codex or Grok session in iTerm2, Terminal, tmux or herdr, with the 2 second hold and Escape. Before it types, brainplane checks that the session is still the one you picked, that it is not at a shell prompt and that the same question is still on screen. Codex sessions also take typed messages through Codex's own queue.

## Stop cards

Sending, paying, deleting, deploying, creating accounts and buying paid plans always stop at a card for you. An agent prepares the draft, fills the form up to the card field or writes the payment recap. You press the final button yourself. No setting moves these actions to automatic.