Canari

Records what a scanner found about an agent component, binds it to the exact artifact bytes it came from, signs it, and lets anyone check the record without trusting us.

What this is, and what it is not. Canari is not a scanner and not a certification. The scanners are third-party tools (bandit, gitleaks, npm audit, and others). Canari records what they reported, binds each verdict to an artifact digest, signs the record, and publishes an append-only log. A verdict means "this tool reported this about these bytes" — never "this component is safe". A graded badge is not a security guarantee, and absence of a record is not evidence of anything.

Check something right now

I want to…Go to
Verify a record in my browser, with no install and no network/verify/
See the raw scan output for 20 PyPI MCP servers, with the re-derivation script/scan/
Read the append-only ledger and its current head/log/
See the measurements taken against this ecosystem/research/

Use it

# gate an MCP server before it starts, from an existing client config
pip install canari-mcp==0.1.4
canari-mcp-gate --store /path/to/store -- npx -y @modelcontextprotocol/server-filesystem /tmp

# verify a record offline, given the published public key
canari-verify record.bundle.json --key public_key.json

The gate wraps a stdio MCP entry: it resolves the component in the signed ledger, refuses to start it if the ledger blocks it, and otherwise proxies stdio byte for byte so the protocol is untouched. It is a configuration-level gate — whoever can edit the config can remove it. It is not a policy engine and not tamper-proof.

Why a signature and not a score

A score asks you to trust the scorer. A signature asks you to trust arithmetic. Canari publishes the key, the exact bytes that were hashed, and a verifier that runs offline, so a third party can disagree with our verdict while still confirming the record is the one we signed and has not changed since.

Limits, stated up front