Docs · MCP tools

The agent drives the manager.

Some facts only the agent has: what the session is really about, which repo it moved into, which branch the work merges into. Rather than scrape that out of the pane, the manager ships an MCP server in the same binary (agent-manager mcp, stdio) and registers it into every session it spawns or revives, so these arrive as native tools with descriptions telling the agent when to call each.

The tools

• renameNames its own session once it knows what the work is, replacing the placeholder.
• review_repoDeclares the repo or worktree it moved into. The path is validated as a git root, so review opens on the right one.
• review_baseDeclares the ref the branch diffs against, validated in that repo and stored per session and repo. Clearing it returns to automatic detection.
• review_modeSets the scope review opens on, for example switching to staged just before it commits.

A stored ref that stops resolving surfaces as an error in review, and B opens a target picker, the repo's branches plus an auto entry, to set or clear it by hand.

Registration

Registration is per tool and automatic. The built-in claude, codex, opencode, grok and gemini tools register on their own:

• claudeA generated --mcp-config file.
• codex-c mcp_servers… overrides on the command line.
• opencodeAn OPENCODE_CONFIG merge file.
• grok, geminiA one-time mcp add --scope user entry on first launch.

A custom tool opts in with mcp = "<style>" in its config block, or out with mcp = "none". There is no prompt injection and no per-project setup. One stdio server serves whichever session called it, identified through its environment, so there is nothing per-session to launch or clean up.

The rename, review-repo and review-base CLI subcommands keep working everywhere, MCP or not, so an agent that speaks no MCP can still run one shell command.