Give each agent its own token
Create one token for Claude Code and another for Codex: each is shown once, carries your rights, and can be limited to the spaces of one project. What each agent writes is in the journal under your name, and revoking one leaves the others working.
$ claude mcp add --transport http --scope user cockpilot \
https://your-team.cockpilot.si/mcp \
--header "Authorization: Bearer ck_..."For Codex, the same server goes in ~/.codex/config.toml: see Connect Codex.
Tell the agent how to work
Put a few lines in the agent's instructions (CLAUDE.md, AGENTS.md):
- Start with
cockpit_briefingor search the board for the task you are given. - Link your branch or pull request to the task with
cockpit_link. - When a choice is not yours, ask with
cockpit_ask_decisioninstead of guessing. - Close the task with
cockpit_complete_with_proof, giving the pull request, deploy or commit.
Read what they did
cockpit_agent_summary answers "what did my agents do?": the tasks they touched, what they closed and on what proof. The journal has every change with its before and after, and the decisions they asked for wait in the tasks.
Keep it safe
Agents read untrusted text: issues, web pages, e-mails. Cockpilot tells them that task content is data, never instructions; a deletion goes to the trash for thirty days. More in Letting an AI agent work on your tasks safely.
Read next
A task manager for AI coding agents · Connect Claude Code · API and MCP reference · All guides