Repository navigation
Gatekeeper 0.2.1
First release. Gatekeeper decides which Claude Code agent or skill should handle a prompt before Claude picks one: YAML rules decide what code can decide, and Jev (TypeSafe System One) makes the one semantic call as a typed Choice with a confidence.
Highlights
- Two rosters. Route to your installed agents, skills and plugins, or to Markdown procedures listed in a JSON index.
- Rollup by group. When no single skill is confident, probabilities are summed by role, so a vague request still reaches the right team.
- Three modes.
shadowlogs what it would do,adviseadds the route as context and never blocks,enforcealso denies calls that break a confident route or aforbidrule. Every install starts in shadow. - Fails open. A Jev outage, a timeout, a missing dependency or a bad argument never blocks a session.
doctor. Catches the silent failures: a missing key, a retired model, hooks or skill links left behind by a moved folder, a rising fail-open rate.- Example gate.
examples/marketing-team/: 24 skills in 11 roles, advise mode, rollup and a bench.
Security and privacy
- Your TypeSafe key is read from the environment,
GATEKEEPER_ENV_FILEor~/.config/gatekeeper/env, and is never printed or logged. - Prompts are masked (the key itself, API keys, tokens, JWTs, private keys, credentials in URLs) before they go to TypeSafe or the local log. The log keeps a hash and a 120-character excerpt, readable only by you.
- See SECURITY.md for what is sent and how to report a problem.
Install
git clone https://github.com/AgriciDaniel/gatekeeper.git && cd gatekeeper
pip install -e ".[test]" && python3 -m pytest -q
python3 scripts/install_hooks.py --global --skill # dry run
python3 scripts/install_hooks.py --global --skill --apply # shadow mode
bin/gatekeeper doctorChecks
166 offline tests (93% line coverage) on Python 3.11, 3.12 and 3.13; clean-container install, hook and uninstall tests; a security audit and a docs-versus-code audit. Full details in CHANGELOG.md.