Repository navigation
Releases
Releases · WaelAbouceo/vouch
v0.5.1 — See what your AI agents can actually do
Vouch is the trust layer for AI agents. One command audits every skill installed on your machine, tells you in plain English what each one can do, and flags the risky ones — deterministically, with zero false alarms.
pip install vouch-agent # zero dependencies; static analysis works out of the box
vouch --audit # scan every skill on this machine
# or, zero-install:
pipx run --spec vouch-agent vouch --auditHighlights of the 0.5 line
- 🔎 Machine-wide audit (
vouch --audit) — auto-discovers every skill for Claude, Cursor, Codex & friends, classifies each, names what it behaves like (Data Courier, Remote Code Runner, File Editor, Advisor…), and shows which need a look. Saves a baseline and reports what changed on later runs. - 🧠 Plain-English profiles — every skill gets a "what this means for you" and a role, inferred from executable context (code/scripts, not prose).
- 🤝 Provider-agnostic AI review (optional) — layer any OpenAI-compatible endpoint, SovereignEG, or the Cursor SDK on top with
--llm. The LLM can only flag for review — it never brands a skill "malicious" on its own, so the strongest verdict stays deterministic and reproducible. - 📊 Labeled benchmark (
bench/) — measure it yourself: 100% precision (zero false accusations) on "malicious", 91%/91% on "flag-for-review", rising to 100% flag-for-review recall with the AI layer.
Fixed in 0.5.1
- GitHub Action install — the Action installed a non-existent
vouchpackage; it now correctly installsvouch-agent, so CI usage works. - Added
vouch --version. - Ship the
py.typedmarker so downstream type checkers see Vouch's types. - Real author metadata; doc bumps.
Full details in the CHANGELOG.
Vouch v0.3.0
First public release of Vouch — the trust layer for AI agents.
Highlights
- Capability-composition gate: verdicts are weighted by capability class, not just findings. A Skill with a dangerous combination (network + credentials, or network + shell) can't reach a clean
validfrom a static-only scan — it's floored tosuspiciousuntil an LLM pass or human--sign-offclears it. - Agent CV: aggregate trust profile across every Skill an agent loads (
--agent-cv). - Skill CV: one-page profile card per Skill (
--cv). - Four delivery forms: Python library,
vouchCLI, MCP server, and HTTP API. - Animated demo GIF in the README.
See CHANGELOG.md for the full list.
Pre-commit users: pin rev: v0.3.0.