The difference from thinking out loud in prose is that a change of mind is structured: `isRevision` with `revisesThought` names exactly which step is being replaced, and a branch keeps its own id, so a chain that doubled back can still be read afterwards. The five metrics are the server's own scoring of each thought — useful as a relative signal across a chain rather than as an absolute measure of the reasoning.
A single-tool MCP server for step-by-step reasoning. Each call carries one thought with its position in a planned sequence, and the response comes back with per-thought metrics and running analytics across the chain, plus a rendered box view of the current step.
- A thought recorded with its number, the total expected, and whether another is needed — the total is adjustable mid-chain via `needsMoreThoughts` — `rat`
- Revision of an earlier step: mark the thought `isRevision` and name the thought it replaces with `revisesThought`, instead of quietly contradicting it
- Branching: set `branchFromThought` and a `branchId` to explore an alternative line without abandoning the first
- Five scores on each thought — complexity, depth, quality, impact and confidence — plus chain-level analytics including total thoughts, average quality and chain effectiveness
- A `visual_output` rendering of the current step and its metrics, for clients that display tool results verbatim
No account and no key. The package.json names `mcp-server-rat-node` (0.1.0), which npm does not have; install globally from a clone with `npm install -g .`, or directly from the git URL, then have the client launch the `mcp-server-rat-node` command it puts on your path. Verify before wiring it up: running it with `--help` should print that the server is running on stdio.
One command — npx -y github:stat-guy/retrieval-augmented-thinking