Gatekeeper

Checked, not fully reviewed. We ran its 168 tests (all pass, offline) and drove the engine with the bundled marketing-team rulebook against a stand-in we pointed its client at. No Jev calls.

A gate is a YAML rulebook. Slash commands, keywords and field checks are settled in code; whatever is left becomes one Jev choice over the skills still in play, and the verdict is one of allow, route, suggest, confirm, escalate or block. When no single skill is confident but the likely ones belong to the same role, it adds their probabilities and routes to the role instead — a sensible use of the distribution Jev returns. A fresh install runs in shadow mode, logging what it would have done and changing nothing.

What leaves the machine is the prompt, and only the prompt. We planted four secrets in one:

planted	sent
an OpenAI-style sk-proj-… key	no — [redacted]
an AWS access key id	no
password=…	no
“the staging db pass is mauve-otter-77”	yes

That is what its own docstring promises: “best effort, not a guarantee” for common token shapes. The TypeSafe key itself is removed from every error message.

It behaves well when things go wrong. A 529 was retried and the prompt still routed; with the endpoint unreachable, both shipped rulebooks fail open — the prompt goes through and the error is logged — so a Jev outage never stalls a session. The rulebook pins jev-1.13.0 on purpose, with a comment saying thresholds are tuned per model version. One gap that only matters against a misbehaving server: a choice outside the offered skills was routed as given rather than refused.

MIT, 19 stars, Python 3.11+.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 6, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.