jev-mcp
Unofficial, read-only MCP server exposing Jev as four advisory tools: decide, route, risk score and yes/no gate.
Overview
jev-mcp is a community-maintained MCP server that exposes TypeSafe AI's Jev as four conservative, read-only tools for bounded probabilistic decisions. It is designed as a second-opinion layer for coding agents, not a replacement for the host model. Typical uses are choosing between retry, rollback and changing strategy, routing to a known tool, estimating change risk, or checking a yes/no gate against a threshold.
The README sets a clear precedence: deterministic evidence such as tests and compilers beats the host model's reasoning, which beats Jev's advice. Every result carries metadata marking it as secondary advisory and not to be treated as fact or as authorisation for destructive work. The tools never edit files, run commands or change the user's model. The API key stays in the local environment or a gitignored .env, and the README warns that anything in state is sent to TypeSafe.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | jev_decide: which of 2-255 explicit alternatives is best? | — |
| Choice | jev_route: which known tool, subsystem or workflow should handle this? | — |
| Score | jev_risk_score: what is the change risk? | low · medium · high · critical |
| Yes / no | jev_gate: does this yes/no condition hold above the threshold? | yes · no |
How it works
- Host agent gathers deterministic evidence first
- It calls a jev-mcp tool with minimal, redacted state
- Jev returns a probability signal wrapped in advisory metadata
- The host model makes the final decision, with evidence overriding Jev
Strengths
- Read-only tool surface cannot modify files or run commands
- Explicit advisory metadata on every result
- Key kept out of MCP client configuration
- CI and security documentation included
Caveats
- Unofficial and not affiliated with TypeSafe
- Advisory only; not meant to authorise destructive actions
- State is sent to the TypeSafe API
Getting started
See the repository's one-click install and quick start docs; requires Node and a TypeSafe API key in the local environment or .env.At a glance
- Type
- MCP server
- Status
- beta
- Language
- TypeScript
- Runtime
- Node MCP server (stdio)
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
\`jev-mcp\` exposes TypeSafe AI's Jev decision model as four conservative, read-only MCP tools for \*\*bounded probabilistic decisions\*\*.
logicrw/awesome-jev-projectsUnofficial conservative MCP server for TypeSafe AI Jev.
wh000wh000/awesome-jev-live
Frequently asked questions
What decisions can it make?
Choosing among explicit alternatives, routing to a known tool, a low-to-critical risk score, and a yes/no gate against a threshold.
Can it approve destructive actions?
No. Results are advisory and explicitly marked as not authorising destructive or irreversible work.
Is it official?
No. It is an unofficial community project not affiliated with TypeSafe AI.
Related projects
Claude Code hook that decides which agent or skill should handle a request using YAML rulebooks plus one Jev Choice with confidence bands.
HookbetaPythonTwo stdlib-only Hermes Agent plugins: a Jev pre-tool gate with a jev_ask tool, and a skill router that names one skill per turn.
Agent skillbetaPythonAgent toolkit for Claude Code and Codex whose optional /d router uses Jev to pick agents, skills and workflows and gate dispatch on intent.
ProjectbetaPythonSkill and CLI that compares a shortlist of agents, MCP servers or skills against a task and permission ceiling, abstaining on weak leads.
MCP serverprototypeTypeScriptComposio agent SDK with an optional TypeSafe provider that uses Jev to shortlist and pick tools and closed-set arguments.
ProjectproductionTypeScriptDeepSeek Harness plugin that registers a native jev_decide tool and declares its own policy for when the agent should ask Jev.
PluginbetaJavaScript