Typed probabilities beat prose for agent decision-making, if the model is calibrated

The core problem this addresses is real and underappreciated: when an agent needs a judgment call, it typically asks a language model, receives a paragraph of hedged prose, and then has to infer a decision from that prose. "This seems fairly urgent" is not a value you can compare against a threshold. The agent has no way to distinguish a confident answer from a guess, and the whole chain becomes fragile.

Typesafe MCP routes those judgment calls to Jev, a model built specifically to return probabilities rather than text. You send a state description and a set of named questions - each typed as a yes/no (noul), a categorical choice, or a numeric score - and get back a JSON object with a probability for every possible answer. The billing/technical/sales example in the README is instructive: the model returns 0.86 for billing, 0.14 for technical, 0.0 for sales, and your agent can act on that distribution directly. A 0.95 urgency score and a 0.55 urgency score are genuinely different signals, and the agent can be wired to escalate the latter rather than treat both as "yes".

A few design choices are worth noting. Batch evaluation is built in: you can pass up to 500 records as items and run the same question set across all of them in a single call, with partial failure isolation so one bad record doesn't abort the rest. The server also ships guidance to the agent on how to write good questions - narrow judgments, structured state, evidence over conclusions - which is a sensible acknowledgment that prompt quality matters even for a typed-output model. Latency is claimed to be under half a second typically, which makes it plausible to call often rather than sparingly.

The distribution story is clean. It's a single static Go binary with no Node or Python runtime, installs via a shell script, and evaluate setup mcp auto-discovers Claude Code, Claude Desktop, and Codex and registers the tool with each. OpenRouter is supported as an alternative API route, and you can point it at a local model server by setting a base URL.

The framing of Jev as a "System One" model - a small, fast unit of AI judgment used like a programming primitive - is a coherent design philosophy. Whether the model itself delivers on that promise isn't verifiable from the README alone; the probabilities it returns are only useful if they're well-calibrated, and nothing here speaks to calibration quality. That's the real question for anyone considering wiring this into a production decision path. But the interface design is sound: typed outputs with explicit probabilities are strictly more useful to an agent than prose, and the MCP wrapper makes the integration path short.