Skills
Public Markdown the agent reads first
A Skill is public Markdown the agent reads before it acts. It says when to call which MCP tool, which fields to send, and how to use the result. It does not widen permissions, and it does not call Jev by itself. The files live on the community skills page.
- Instructions, not a new capability
- No key, no call
- Six files for the agent's skills directory
- After the result comes back
- The tools and the primitives
Instructions, not a new capability
The agent already has whatever tools its host allows. The Skill only says when a Jev judgment is the right next read, and which fields that call needs.
It does not widen permissions. It does not call Jev by itself. Without the MCP connection, there is nothing for those instructions to invoke.
No key, no call
Without MCP and a key, the Skill can only describe the workflow. Downloading a Skill does not spend Jev quota.
JEV Hub does not copy the Markdown into this repository. Each file stays at its source under jevai.org/skills.
Six files for the agent's skills directory
Install the six files in the agent's skills directory. Read them from the community site. Do not paste a key into any of them.
jev
The router. It picks which workflow Skill fits, or whether to call jev_decide. If the six tools are missing, do not invent a successful call. Source: jev/SKILL.md.
jev-tool-guard
Calls jev_guard_tool_call. allow means call inside the scope already judged. confirm means get an explicit yes from the user. review means inspect, then judge once more. deny means do not call. Do not use the result to skip an approval the user or the platform already required. Source: jev-tool-guard/SKILL.md.
jev-model-router
Calls jev_route_model. Use a returned model id only when that model is actually available in the current environment. Source: jev-model-router/SKILL.md.
jev-task-router
Calls jev_route_task. block means stop. It does not mean find another way around the stop. Source: jev-task-router/SKILL.md.
jev-research-guard
Calls jev_check_research. Jev does not browse and does not fetch new sources. Source: jev-research-guard/SKILL.md.
jev-completion-review
Calls jev_review_completion. incomplete means the goal is not finished. Source: jev-completion-review/SKILL.md.
After the result comes back
Follow the returned decision unless it conflicts with a higher-priority instruction or a safety policy. A probability is not permission to skip a human approval.
If the facts, the parameters, the candidates, or the scope change, call again. Do not send passwords, API keys, or private data the judgment does not need.
The tools and the primitives
The calls those files name are the six remote jev tools. The boundary for the agent is a judgment at the decision point.
Question types for jev_decide are Choice, Score, and Noul. The request builder on this hub validates JSON and does not call an API.