Jev-AI-Skill
One skill and MCP server for Claude Code, Codex and Hermes where Jev gates large-model turns, routes models and guards Git steps.
Overview
Jev-AI-Skill is a single skill plus MCP server for Claude Code, Codex and Hermes that keeps large models out of turns a cheap decision can settle. A Loop capability turns a build or repair goal into steps, runs configured checks and requests an independent review, saving state for resuming. A Gate capability asks Jev before a large model runs, through four MCP tools: gate_triage (does a cron run, poll, notification or chat event need an agent at all), gate_verdict (which review findings the evidence already closes), gate_decide (answers a question that would otherwise go to the owner, and says when the owner really must decide), and gate_pick (chooses a tool, skill or session from one-line descriptions).
Route picks a model tier, a skill and whether to delegate; a Git capability recommends commit, push or pull-request steps within repository rules; and find_code.py ranks code-search results so the coding model reads less. A script-only watch mode runs recurring checks with no agent, reporting new failures once, and an HTTP mode keeps the TypeSafe key on one server for a private network. It has a test workflow and an MIT licence.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this cron run, poll, notification or chat event need an agent at all? | yes · no |
| Choice | Which tool, skill or session fits this task? | — |
| Yes / no | Does the evidence already close this review finding? | yes · no |
How it works
- Event or goal arrives
- gate_triage decides whether to wake an agent
- Route picks model tier and skill
- Loop runs checks and independent review; gate_verdict closes settled findings
- Git capability recommends the next step
Strengths
- Avoids large-model turns for routine events
- Central HTTP mode keeps the key on one server
- Watch mode reports new failures once with no agent
Caveats
- Broad scope; several moving parts to configure
- Requires a TypeSafe key
At a glance
- Type
- MCP server
- Status
- beta
- Language
- Python
- Runtime
- Agent skill, MCP server
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
One AI skill + MCP server for Claude Code, Codex and Hermes: Jev (TypeSafe) gates large-model turns (event triage, review verdicts, owner questions, tool choice), runs build-and-repair loops with independent review, routes models and skills, and guards Git steps. Script-only watch mode and a shared HTTP server.
logicrw/awesome-jev-projectsOne AI skill + MCP server for Claude Code, Codex and Hermes: Jev (TypeSafe) gates large-model turns (event triage, review verdicts, owner questions, tool choice), runs build-and-repair loops with independent review, routes models and skills, and guards Git steps. Script-only watch mode and a shared HTTP server.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev gate?
Whether events need an agent, which review findings are closed, owner questions and tool choice.
Which agents are supported?
Claude Code, Codex and Hermes.
Can it run without an agent?
Yes, --watch runs recurring checks script-only.
Related projects
Skill pack handing an agent's small decisions to Jev: model routing, skill choice, search, memory, compaction and computer use.
Agent skillbetaPythonSelf-improving multi-platform agent harness with an optional Jev companion for typed judgments and per-session model routing.
ServicebetaTypeScriptFourteen opt-in Claude Code and Codex skills where Jev makes one bounded choice: which skill, model, test, file or next step.
ProjectbetaPythonNode experiment where Jev picks the next step around a coding agent, with approved actions and Codex or Claude Code delegation.
ProjectprototypeTypeScriptClaude Code skill that splits work into PBIs for Codex, Antigravity or Claude sub-agents, using Jev for model choice and triage.
ProjectprototypePythonRuns Claude Code on Opus 5.5 with Jev re-picking reasoning effort after every tool batch, via per-message effort so the cache survives.
ProxybetaTypeScript