What Is Jev? TypeSafe's System One AI Model Explained

Choice selects one option from a finite set and returns probabilities across the alternatives; Score places the state on an ordered rubric defined by the application; Noul asks whether a proposition is true and returns a probability from zero to one.

Its primary output is a typed decision with probabilities rather than text, its answer space is defined by the application, and control flow branches on validated values. Generative models remain better at writing, reasoning, synthesis, and coding.

An independent, community-built MCP server maintained by jkudish, open source under the MIT licence but not an official TypeSafe AI server. TypeSafe provides the Jev model and official SDKs; the MCP integration is a separate project.

It runs locally over stdio and can be connected to Codex, Claude Code, Cursor, OpenCode, and other MCP clients. The current npm package is @jkudish/jev-mcp, run with npx -y @jkudish/jev-mcp.

Through TypeSafe's API and supported gateways; the article documents directed TypeSafe access along with Vercel AI Gateway, Cloudflare Workers AI, and OpenRouter paths, and lists typesafe-ai/jev at $0.04 per million input tokens on Vercel as of September 18, 2026.

Construct a minimal state, define a closed answer space or rubric, retain full probabilities and confidence, validate the returned shape in ordinary code, apply thresholds tested on representative data, escalate uncertain or high-impact cases, and keep authorization and irreversible actions outside the model.

Writing articles, producing code, multi-hop research, and deterministic checks. If a value can be calculated exactly, calculate it in code; if an operation needs authorization, enforce it in code; if the answer needs open-ended synthesis, use a generative model.