jev-use plugin
A plugin and MCP server that hand no-text decisions to the Jev model and gate shell commands while the LLM writes content
- 71 ★ on the repository
- Last commit
- MIT
Extended access
Why this level
- The PreToolUse hook sees every agent command and sends its text to the model provider with credentials redacted before sending
- Task state goes to a third-party provider, TypeSafe, OpenRouter or Vercel AI Gateway, under your key
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/jev-useDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewExtended accessThe tool runs code or makes network calls and only reads external data.
- The PreToolUse hook sees every agent command and sends its text to the model provider with credentials redacted before sending
- Task state goes to a third-party provider, TypeSafe, OpenRouter or Vercel AI Gateway, under your key
This is third-party code. Review the repository files before installing.
What it does
jev-use connects the fast Jev judgment model from TypeSafe to Claude Code, Codex and pi. The idea is a division of labor: when text has to be written, the LLM works, and when a step needs only a fast decision, Jev takes it. npx -y jev-use install writes the needed settings into every agent it finds. The server gives two MCP tools. jev_judge takes a state and a batch of questions of three types (noul rates the probability of a statement, choice picks an option, score sets a position on an ordered scale) and returns an answer with confidence. If confidence falls below a threshold the answer is marked escalate and the LLM decides. jev_gate checks one proposed action and answers allow, deny or escalate. An allow stays silent and hands the decision to the normal permission flow, so the gate never grants anything itself and only denies or asks. If the provider is down the result is escalate with the reason unreachable. Credentials in the action are redacted before sending. The plugin also contains a skill with routing rules and a PreToolUse hook. The Jev model connects through TypeSafe directly, OpenRouter or Vercel AI Gateway. JEV_BACKEND=mock gives a keyless dry run, and npx -y jev-use doctor checks the wiring. The same code is available as a library with the check, pick and rate calls.
Who it is for. Agent developers and anyone who wants to speed up Claude Code and Codex steps that need a yes or no with no long text, and put a cheap safeguard on shell commands.
Good fit when
- Every shell command needs a fast filter: dangerous commands are denied within hundreds of milliseconds with a reason
- A 200-message context is assessed in a few calls while the LLM writes only the final paragraph
- An agent has to pick the next step from CI state: merge, rerun or hold
- Route steps go through a yes or no check with confidence and doubtful ones go to the LLM
- Application code needs a library with check, pick and rate questions and an escalation contract
Not a fit when
- You need a fully offline setup: judgments go to a provider, and only the JEV_BACKEND=mock mode is local
- You need a gate that approves actions itself: allow only passes the decision to the normal permissions
- You need long coherent text: Jev decides short questions and the LLM writes the text
- You have no TypeSafe, OpenRouter or Vercel AI Gateway key: without one only the mock works
Example request
Install jev-use and check every shell command through jev_gate, deny dangerous ones with a reason and take decisions that need no text from JevLimitations
- One key is needed in the agent environment: TYPESAFE_API_KEY, OPENROUTER_API_KEY or AI_GATEWAY_API_KEY. The JEV_BACKEND=mock mode stays local.
- The judged state goes to the chosen provider, and credentials are redacted from the checked action before sending.
- The allowed state size is about 30 thousand tokens per call.
- The gate never grants permissions itself: on allow the normal permission order applies, and if the provider is unreachable the result is escalate.
- Latencies in the README were measured on the author micro-benchmarks, and your key and region will give other numbers.
- The project is largely written with Claude Code, as the README says. The license is MIT.
How to disable. Connections are removed by deleting the jev-use entries from Claude Code, Codex and pi settings and removing the plugin. The package in a project is removed with npm uninstall jev-use. npx -y jev-use doctor shows what is left.
FAQ
How do I install jev-use?
With npx -y jev-use install: it wires Claude Code, Codex and pi, whichever it finds. The wiring is checked with npx -y jev-use doctor.
Which keys are needed?
One provider key: TYPESAFE_API_KEY, OPENROUTER_API_KEY or AI_GATEWAY_API_KEY. Without a key you can try the JEV_BACKEND=mock stub.
What does jev_gate do?
It checks one agent action and answers allow, deny or escalate. An allow stays silent, so the decision stays with the normal permissions, and with an unreachable provider the result is escalate.
What is escalation in jev_judge?
If Jev confidence is below a threshold (0.5 for reported and 0.4 for estimated confidence) the answer is marked escalate with a typed reason, and the LLM decides the question.
Can I use jev-use as a library?
Yes, npm i jev-use. The Jev class is called with the judge method and check, pick and rate questions, and returns answers with confidence and an escalate flag.
Discussion
No questions or reviews yet.
Related
PluginsA plugin with a catalogue of public Jev examples: a skill and an MCP server show how others solved the same decision task.
An MCP server giving an agent three Jev tools: typed yes or no, choice and score questions with probabilities, offline by default.
Jev-Style runs small decision models on your machine: a probability per option, six skills, MCP and a Claude Code guard hook
The Jev classifier as an agent tool: labels items, answers yes or no and ranks candidates with probabilities within milliseconds.
An MCP and ACP adapter for the TypeSafe Jev decision model: typed answers and confidence gates for clicks and tool calls.
A local router merges models, subagents, skills and MCP tools into one candidate set and picks who handles the next step.
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| jev-use (this card) | no | key is optional | 71 | |
| JevRouter | no | 469 | a request router built on Jev with no command gate or MCP tools | |
| awesome-jev | no | not required | 602 | a collection of Jev material with no installable plugin |
README in short
The README presents jev-use as a way for Claude Code, Codex and pi to work with the Jev model: fast steps with no text go to it, and content stays with the LLM. It shows demos with measurements, installation and provider keys, library use, a file map and development commands.