jev-agent-draft
Skill and CLI that compares a shortlist of agents, MCP servers or skills against a task and permission ceiling, abstaining on weak leads.
Overview
Jev Agent Draft helps choose which agent, MCP server, skill or plugin to trust with a task before granting access. It compares a finite shortlist against an explicit task and a permission ceiling. Hard gates run first: requested permissions must be within the allowed set, maintenance age must be under a limit, every candidate needs evidence references, and every dimension needs a finite 0-100 rating. Eligible candidates are then weighted by task fit, permission minimality, evidence quality, maintenance and portability, and the leader must clear both a minimum score and a minimum margin or the draft abstains.
Each result is emitted as a tamper-evident Jev Decision Receipt. The demo picks the strongest eligible code reviewer and rejects a capable agent that asks for repository write access, with no dependencies, key or network calls. It is packaged as a Codex/ChatGPT plugin where skills teach the comparison workflow while a deterministic local CLI owns scoring, and it has a test workflow.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate agent is fit for this task under the permission ceiling? | — |
| Other | Does the leader clear minimum score and margin, or should the draft abstain? | select · abstain |
How it works
- List candidates with permissions, maintenance, evidence and ratings
- Apply hard gates
- Weight eligible candidates across five dimensions
- Select the leader or abstain if score or margin is too low
- Write a tamper-evident decision receipt
Strengths
- Excess permissions are hard failures, not footnotes
- Abstains rather than giving false certainty
- Demo runs offline with no key
- Receipts make decisions auditable
Caveats
- Ratings are supplied inputs; quality depends on them
- Jev's role versus deterministic scoring is limited in the default CLI
Getting started
`git clone https://github.com/gbesse/jev-agent-draft.git && cd jev-agent-draft && npm test && npm run demo`; or `codex plugin marketplace add gbesse/jev-agent-draft --ref main`.At a glance
- Type
- MCP server
- Status
- prototype
- Language
- TypeScript
- Runtime
- CLI, Codex plugin
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Choose the right agent, MCP server, or skill before granting access.
wh000wh000/awesome-jev-live
Frequently asked questions
What does Agent Draft decide?
Which candidate agent, MCP server or skill is fit for a task within your permission limits, or to abstain.
Does the demo need an API key?
No. It runs with no dependency, key or network request.
What is a decision receipt?
A tamper-evident record of the decision, in the Jev Decision Receipt format.
Related projects
Agent toolkit for Claude Code and Codex whose optional /d router uses Jev to pick agents, skills and workflows and gate dispatch on intent.
ProjectbetaPythonClaude 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 skillbetaPythonOpen-source Jev skills for Claude Code, Codex, Cursor and Grok: one client to gate, route, guard, verify, triage, review and act.
ProjectbetaTypeScriptLocal gateway and MCP tools that let Jev predict which tool a coding agent should call next, from suggestions to enforced choices.
MCP serverbetaTypeScriptUnofficial, read-only MCP server exposing Jev as four advisory tools: decide, route, risk score and yes/no gate.
MCP serverbetaTypeScript