Install ChangePilot in your connected agent

Stored in ~/.claude.json. Restart your Claude Code session to pick up the new server. This form is anonymous: the tools appear, and planning works — but calls that need your account (assignments, history, proposals) need a token. Add it with the colon-syntax header:

Easiest way to get the token: the connected-agent flow above (no copy-paste), or mint one at account settings.

Recommended (desktop): install the session plugin so ChangePilot engages on every session (hooks + device-flow connect). Fetch /api/install-mcp and follow session_autostart.clients.cursor.runbook — or run:

MCP-only (no session autostart): add to .cursor/mcp.json (project) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "changepilot": {
      "url": "https://changepilot.deltamarksystems.com/api/mcp/v1",
      "type": "http",
      "headers": {
        "Authorization": "Bearer mcp_pat_..."
      }
    }
  }
}

Restart Cursor to pick up the new server.

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "changepilot": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://changepilot.deltamarksystems.com/api/mcp/v1"]
    }
  }
}

Config file location: macOS ~/Library/Application Support/Claude/, Windows %APPDATA%/Claude/. Restart Claude Desktop.

Add to ~/.codex/config.toml under [mcp_servers]. Codex uses the env var name in bearer_token_env_var, not the token value:

[mcp_servers.changepilot]
url = "https://changepilot.deltamarksystems.com/api/mcp/v1"
bearer_token_env_var = "CHANGEPILOT_CODEX_PAT"
enabled = true

Export the PAT before starting Codex: export CHANGEPILOT_CODEX_PAT=mcp_pat_... — mint via /connect-agent or /connect.

ChangePilot also installs user-level ~/.codex/hooks.json lifecycle hooks. After install, restart Codex, open /hooks, review the two ChangePilot commands, and trust them. The hooks use only the dedicated Codex PAT; they never fall back to a Claude or Cursor credential.

Use Codex tool discovery to load and call connection_health inside Codex; raw HTTP success alone does not verify that the host can use the MCP. If Codex reports 'user cancelled MCP tool call', approve the tool in an interactive session, or for headless codex exec use --dangerously-bypass-approvals-and-sandbox only inside a trusted external sandbox.