FEATURE GUIDE · LOCAL MCP
Choice, Score, and Yes/No MCP Decisions
Ask one to eight bounded Choice, Score, or yes/no questions in a single MCP request and inspect each answer's probability, confidence source, and calibration status.
The API surface is designed for branches in ordinary software: define the possible answers first, ask a small question about supplied state, validate the shape, then decide in code what a score or probability permits.
HOW IT WORKS
From current state to a bounded action
- Choice selects one key from two to 255 named alternatives.
- Score returns a numeric expected position on an ordered rubric of up to ten levels.
- Yes/no (called `noul` in the request schema) represents a judgment as a probability from zero to one.
- Batch independent questions in one call; each result contains its provider, model, latency, confidence source, and calibration status.
- Call `provider_warmup` in the same MCP session before a latency-sensitive run; first use may download and initialize the local model.
LIMITS TO KEEP IN VIEW
The default MiniLM embedding provider is not a generative reasoning model. Its probability distribution is an uncalibrated softmax over embedding similarities. Measure calibration on held-out labels before thresholding it for consequential decisions.