beamWidth and numSimulations are exposed to the model, so how many paths to keep and how many simulations to run become part of the reasoning rather than fixed config. Be clear-eyed about the state of it: the alpha policy-simulation variants are described by the author as incomplete, and the benchmark section is a plan, not results.
A reasoning server for Claude Desktop that runs a problem through an explicit search strategy and reports on the paths it explored.
- Two search strategies you switch between: beam search, and Monte Carlo Tree Search for harder problems
- Two experimental MCTS variants added in v2.0.0 — mcts-002-alpha, which uses A* search with an early policy simulation layer, and mcts-002alt-alpha, which uses bidirectional search
- Scoring across reasoning paths, and a map of the different routes taken through a problem
- An analysis of how the reasoning process went
- Model control over the search parameters: beamWidth from 1 to 10, and numSimulations from 1 to 150
Node and npm. Clone the repository, npm install, npm run build, then point Claude Desktop at dist/index.js. The alpha simulators are explicitly incomplete and subject to change; benchmarks against MATH500, GPQA-Diamond and others are listed as planned rather than published.
Build from source — clone the repository and build it, then point your client at the binary