jev-mcp (Songokou1983)

Self-hosted Jev MCP server for Claude Code and Codex with confidence gating, cost guardrails, caching and decision logs.

MCP serverbeta

View on GitHub ↗

Overview

This jev-mcp is a local MCP server, documented in Chinese, that exposes TypeSafe Jev as native tools for Claude Code, Codex and other MCP clients. It is positioned for high-frequency small decisions inside an agent's reasoning chain, such as routing, classification, scoring and checks, rather than for high-volume production pipelines that should call the API directly.

The author runs it as a systemd user service reachable over Tailscale with bearer-token auth. Tools include classify (2-255 options), check (yes/no with a derived effective confidence), rate (2-10 levels), batch_judge (independent calls per item, capped at 20), a raw System One passthrough and a health probe. Server-side features include confidence gating into act, confirm or escalate, a state size cap and per-session budget, a sha256 LRU cache, metadata-only decision logs, a pinned model version and exponential backoff on rate limits. The README reports 142 passing tests.

What the decision model decides

Type	Question	Answers
Choice	classify: which of the given options fits the text?	—
Yes / no	check: does the text satisfy this condition?	yes · no
Score	rate: which ordered level does the text reach?	—

How it works

Run the server as a systemd user service bound to a Tailscale IP
Clients connect over HTTP with a bearer token
The agent calls classify, check, rate or batch_judge
Results are gated into act (>0.7), confirm (>0.4) or escalate
Decisions are cached and logged as metadata

Strengths

Built-in budget and size guardrails
Caching avoids paying for repeated questions
Pinned model version for reproducibility
Extensive test suite (142 tests reported)

Caveats

Documentation in Chinese and tied to the author's home setup
Model version is hardcoded
Codex client configuration marked TBD

Getting started

git clone the repo, uv pip install -e ".[dev]", put TYPESAFE_API_KEY in ~/.jev/.env and a token in ~/.jev/token, then install the systemd user service.

At a glance

Type: MCP server
Status: beta
Language: Python
Runtime: MCP server (HTTP or stdio)
Has code: yes
Has tests: yes

Use cases and integrations

Infra / SDKs / Integrations537

How curated lists describe it

Local MCP server exposing TypeSafe Jev (System One decision model) as native Claude Code / Codex tools
logicrw/awesome-jev-projects
Local MCP server exposing TypeSafe Jev (System One decision model) as native Claude Code / Codex tools.
hellogumbo/awesome-jev

Frequently asked questions

How are results gated?

By default, confidence above 0.7 is act, above 0.4 is confirm and below 0.4 is escalate.

How does it limit cost?

It caps state size, sets a per-session budget and caches repeated questions.

Can I run it locally without Tailscale?

Yes. A stdio mode is documented for same-machine development.

Related projects

jev-netlify-mcp
Key-locked Netlify proxy that serves Jev through Netlify AI Gateway credits, plus an MCP server for Claude Code, Codex and Claude Desktop.
MCP serverbetaPython
jevify (lucioamor)
Skill and MCP service that audits an app for AI calls that are really decisions and shadow-tests migrating them to Jev.
MCP serverbetaTypeScript
jev-architect (BenjaminPolge)
Skill that makes Claude Code and Codex split AI features into Jev, LLM or plain-code steps before writing code.
Projectbeta
jevkit
Rust CLI for Jev that lints typed question sets offline, then sends them and prints probabilities, labels and scores as JSON.
CLIbetaRust
jevskill
Agent skill that has Jev pick the few relevant lines from big logs before your coding agent reads them, with a reversible reduce and ledger.
ProjectbetaPython
Oko
Local code search for coding agents: ripgrep/BM25 shortlist, then Jev judges each chunk's relevance, served over MCP.
MCP serverbetaRust