Tori's own MCP server

tori mcp is a stdio MCP server that exposes Tori itself as a set of tools: listing sessions, reading a transcript, starting a worktree, creating a checkpoint diff, asking you a question and waiting for the answer, opening a pull request. Every tool call becomes one call on the same protocol the tori CLI and the phone app use, so there is one source of truth for what Tori can do, not three.

How a session gets it

A Claude session gets it automatically: Tori writes a small MCP config file naming tori mcp as a server and passes it on launch. An ACP agent (Codex, OpenCode today) gets it the same way, sent as part of its own session setup, only for adapters measured actually calling a server they were handed. A worker session autopilot spawned gets the same tools, with outward actions (opening a pull request, submitting a review, merging) gated behind an approval you gave beforehand rather than a permission prompt mid-task, since nobody is watching a background worker’s chat directly.

What it will not do without you

Anything that leaves the machine, a pushed branch, a posted review, a merge, needs an approval reserved for that exact action and exact draft content ahead of time. A worker cannot invent its own approval, spend one meant for something else, or reuse one twice.

Reading a session’s tools

Every tool an agent can see is scoped to what kind of caller it is; a worker session, for instance, never sees the tools that would let it spawn another session or start or stop autopilot itself. From inside Tori, a session’s overflow menu > Session shows the live MCP servers it has connected, including this one, and how many tools each exposes.

The CLI

tori mcp is also a subcommand of the plain tori CLI, alongside commands for listing sessions, tailing a transcript, creating a worktree, and the rest of the same tool surface from a shell instead of an agent. Run tori mcp yourself to start the server directly, useful for wiring a session up by hand outside of Tori’s own launch flow.