canary
JoyTiefling/canary
Trust verdict for open-source repos & bounties before you engage — ENGAGE / CAUTION / AVOID / UNKNOWN. CLI + MCP server for agents. Never says "go" on missing data.
Instalação
npx convez-ai JoyTiefling/canary1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
🐤 Canary
Should I — or my agent — engage with this repo/bounty, or is it a trap?
Canary gives a fast trust verdict (ENGAGE / CAUTION / AVOID / UNKNOWN) for a
GitHub repository or bounty before you sink hours (or tokens) into it. It's built
for the open-source contributor and for the autonomous coding agent that now picks
its own work — neither of whom can trust surface signals anymore (stars are faked,
bounties go unpaid, popular issues are swarmed, honeypots target agents).
python -m canary facebook/react
# [ENGAGE] (risk 0.05, confidence 0.85)
# a swarmed bounty looks like this (illustrative output; use a real owner/repo#issue):
python -m canary <owner>/<repo>#<issue>
# [AVOID] (risk 0.97, confidence 1.00)
# • veto: contention — many open attempts, no payout yet (swarmed bounty)
python -m canary --json <owner>/<repo>#<issue> # machine-readable, for agents
What it checks (risk dimensions, not one blended number)
- authenticity — fake-star proxy (fork:star), fast-growth anomalies, releases, owner account age. (Reuses the Dagster/StarScout fork:star heuristic — not reinvented.)
- responsiveness — is the maintainer actually alive (recent pushes)?
- honeypot — agent-bait labels ("Autonomous Agents Only", "crypto-eligible") + bot/throwaway authors.
- availability — already assigned, or reserved (e.g. hiring-only) bounties.
- contention — how swarmed is this bounty? (Algora
/attemptdensity today; generalizing to native open-PR/claimant counts on any issue.)
The one rule that matters
Canary never says "go" on missing data. If it can't gather enough signal, the
verdict is UNKNOWN, not ENGAGE. Absence of evidence is not evidence of safety —
a false green sends you into a trap. (The #1 lesson from validating the approach.)
Use as an MCP server (for agents)
Expose Canary as a tool an autonomous agent calls before engaging a repo/bounty:
pip install mcp # only needed for the server transport
python -m canary.mcp_server # stdio
Tool: canary_check(target, verbose=False) → {verdict, engage, risk, confidence, reasons}.
engage is the go/no-go boolean; UNKNOWN means "not cleared", never treat as safe.
This is the point of Canary: a pre-engagement trust gate for the autonomous agents
that now pick their own work and can't tell a real bounty from a swarm or a honeypot.
Architecture
Platform-agnostic core (universal GitHub signals) + pluggable context modules. Algora is the first bounty module; the design is not bound to it (Opire, IssueHunt, and non-bounty "should my agent touch this repo at all" are next).
Status
Early MVP. Signal set validated on a small labeled sample (promising, not proof — expanding the test set). CLI and MCP server both work today. Roadmap: harden Algora extraction · native contention fallback via linked PRs · larger labeled benchmark · more platform modules (Opire, IssueHunt).
Install / run
No third-party deps. Python 3.10+. python -m canary <target>.
Set GITHUB_TOKEN to raise the API rate limit.
Built by JoyTiefling. Contributions and skepticism welcome.