DEMO / RECORDING
Overview & result
JCR gives a Claude or Codex agent one MCP tool to find documented operations. Jev classifies the request and searches progressively narrower capability-tree branches; a separate model decomposes compound requests. The repo includes a demo and nuanced benchmark receipts.
From the original post
JCR gives an agent one tool to find the documented commands it needs for a task.
How Jev fits in the loop
- Decide with Jev whether the task needs one external action or several; decompose compound requests separately.
- At each tree node, ask Jev to rank direct child capabilities, including a no-match option.
- Return selected command documentation to the agent; execution remains outside JCR.
How to reproduce
- Clone JCR, run `npm ci` and `npm run capabilities:audit`.
- Configure TypeSafe and a chosen agent key, then resolve a task such as the README's Stripe example.
- Run the documented skills-vs-JCR harness; compare context, cost, latency and the lookup-only task boundary.
Why this build matters
Shows Jev as a skill/capability retrieval layer and documents a real tradeoff: less context can come with higher end-to-end latency.
Limitations
- JCR returns documented commands; it does not execute them, and its benchmark harnesses also stop at explaining steps.
- The author's benchmark used one run per scenario/variant; Codex timing worsened in most tested scenarios.