Jevkit — maxivanyshen

A self-hosted set of Jev tools for coding agents: a small HTTP server that holds the TypeSafe key, stdlib CLIs called ask-jev and jev-review, agent skills and a Pi tool. Agents ask Jev typed questions about code instead of reading it into their own context.

Either docker build -t jevkit . and docker run -e TYPESAFE_API_KEY=... -p 8787:8787 jevkit, or, without Docker, TYPESAFE_API_KEY=... python3 -m jevkit.server. PORT defaults to 8787 and the key comes from console.typesafe.ai.

It evaluates a state — or a batch of up to 100 items — against typed questions: noul for a yes/no probability, choice for an option plus distribution, and score for a probability-weighted position on a rubric. Responses hold answers keyed by id plus token usage.

The body is a raw diff. It is split per file, a fixed triage rubric is asked (needs_review, risk, category, security_concern, missing_tests), and flagged files come back with their full diff. It also runs without the server: git diff | TYPESAFE_API_KEY=... python3 -m jevkit.review.

No. install.sh puts ask-jev and jev-review into ~/.local/bin, the skills into ~/.claude/skills and ~/.agents/skills when present, and the Pi extension into ~/.pi/agent/extensions; the server is what holds the key, so no client-side key is needed.

Passing private as true answers with Kev-0.8B, open weights on the homelab, instead of TypeSafe. Nothing leaves the homelab, but it runs on CPU at about 2.3 s per item and clips states to 32K characters; KEV_URL defaults to http://kev:8009/v1/systemone.

Auth is optional: unless JEV_ACCESS_KEY is set, the POST endpoints are open to anyone who can reach them. States are clipped at 200 KB, then halved until they fit Jev's 32k-token limit, and truncated items are flagged _state_truncated. This entry reflects the truncated README.