jev-rs
Rust engine that answers noul, choice and score from any LLM in one prefill via logprobs, wire-compatible with /v1/systemone and MCP.
Overview
jev-rs is a Rust engine that produces System One judgments from any language model in one prefill. It answers typed questions about a state (noul, choice, score) with probabilities read from the model rather than generated text, and is wire-compatible with TypeSafe's POST /v1/systemone so TypeSafe SDKs can point at it. It also runs as an MCP server with a single judge tool for Claude Code, Codex, Grok Build and OpenCode.
It describes itself as the harness around the open Jev ecosystem rather than a model or single technique: a logprob scorer for any decoder behind llama-server (any GGUF), adapters for trained specialist models, the TypeSafe-compatible wire server, and one eval and calibrate metric stack to compare them side by side. OpenAI-compatible backends such as the DeepSeek API, vLLM or SGLang work through chat completions with logprobs, though the raw llama.cpp path is exact and preferred. An install script provides prebuilt binaries for macOS and Linux.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team handles this? | billing · technical · sales |
| Score | How urgent? | not urgent · soon · blocking |
| Yes / no | Does the customer threaten to leave? | yes · no |
How it works
- Install jev and start a GGUF model behind llama-server
- Ask questions via CLI, the /v1/systemone server or the MCP judge tool
- jev-rs reads option probabilities in one prefill
- Evaluate and calibrate models with the built-in metrics
Strengths
- Works with any local decoder model
- Drop-in for TypeSafe SDK clients
- MCP tool for coding agents
- Shared evaluation and calibration tooling
Caveats
- Calibration depends on the underlying model
- OpenAI-compatible path depends on first-token behaviour
- Requires running a model server
Getting started
curl -fsSL https://raw.githubusercontent.com/yijunyu/jev-rs/main/install.sh | sh, start llama-server with a GGUF model, then jev --backend http://127.0.0.1:8089 ask ...At a glance
- Type
- Model
- Status
- beta
- Language
- Rust
- Runtime
- CLI, HTTP server and MCP server
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
System One judgments (noul/choice/score) from any LLM in one prefill — a Rust, Jev-compatible /v1/systemone engine
logicrw/awesome-jev-projectsSystem One judgments (noul/choice/score) from any LLM in one prefill — a Rust…
wh000wh000/awesome-jev-liveSystem One judgments (noul/choice/score) from any LLM in one prefill — a Rust, Jev-compatible /v1/systemone engine.
hellogumbo/awesome-jevRust engine that answers the three primitives from any language model in one prefill, reading probabilities rather than generating.
MrJev/awesome-jev
Frequently asked questions
Does it need TypeSafe?
No. It reads probabilities from local or OpenAI-compatible models.
Can TypeSafe SDKs use it?
Yes. It serves a wire-compatible /v1/systemone endpoint.
How do coding agents use it?
Through jev mcp, an MCP server with a judge tool.
Related projects
MCP server exposing Jev-style decision models served locally by Ollama's /v1/systemone as classify, yes/no and score tools.
MCP serverbetaTypeScriptDocker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.
MCP serverbetaPythonNoulXP: open standard and reference engine for packaging calibrated System One decision models so one runtime can run them all.
ProjectbetaPythonMCP server, CLI and Claude Code skill with Choice, Score and Noul primitives and a confidence gatekeeper, hosted or offline.
MCP serverbetaTypeScriptGo CLI and MCP connector exposing an evaluate tool for Jev and other System One models to Claude Code, Claude Desktop, Codex, Hermes and pi.
CLIbetaGoMCP gateway that trims noisy tool output and offers task classification and review advice, splitting work between local Laya and Jev.
MCP serverprototypeTypeScript