I'm the maker of Jevable. This is a technical walkthrough of one shipped example, not a production-readiness or user-count claim.

Coding agents invent steps easily. Recurring work often needs an accepted graph, a retained run ID, and a human decision before anything continues.

Jevable is an open-source (MIT) agent skill on GitHub (0.1.0-alpha.1). An agent can turn a described process into a workflow app you can inspect, test, and run independently. This post covers only the model-free request-review example. No hosted model. No outbound send.

Install:

git clone https://github.com/somethingwentwell/jevable-skill cd jevable-skill npm ci --omit=dev

Needs git, Node.js 22.19+, and npm. The checkout ships prebuilt dist/, so normal use has no build step. Point your agent at that checkout as the skill location.

The example routes on an input boolean called needsReview:

  • needsReview: false → the routine path completes (exit 0)
  • needsReview: true → Human Approval waits (exit 3); do not approve just to “finish” the demo

needsReview is authored in the workflow. It is not automatic sensitivity detection.

You should see one completed run, one waiting approval with retained run IDs, and no automatic approve/reject.

Limits: early alpha — not production-ready; hosted model providers remain unverified in this public package; custom code nodes validate but do not execute; some public test helpers / boundary tooling are reconstructions, not private-baseline byte matches.

If you try the clone path, comment with your OS/Node version and the first error you hit.

Repo: https://github.com/somethingwentwell/jevable-skill

License: MIT

Earlier Cursor Forum thread: https://forum.cursor.com/t/jevable-an-mit-agent-skill-for-runnable-workflows-with-human-approval/177901

Disclosure: I used AI assistance to draft and tighten this article; I reviewed the technical claims and am responsible for the content.