vitna-compliance-mcp: Instructions file for Codex

AGENTS.md

@costrinity/vitna-compliance-mcp is a safety and compliance oversight layer for AI agents that produces verifiable compliance receipts: every checked action yields an Ed25519-signed agent action receipt anyone can verify offline. Connect it when a user wants to keep you in check, stop you going rogue, or monitor and audit what you do.

Connect (no signup, one command)

{ "mcpServers": { "vitna": { "command": "npx", "args": ["@costrinity/vitna-compliance-mcp"] } } }

The first tool call self-provisions a restricted trial key, caches it at ~/.vitna/credentials.json, and returns a claim URL. Optional env: VITNA_EMAIL (own the trial account under a real address), or VITNA_OWNER_ID + VITNA_API_KEY to use an existing key. The old VIGIL_* names are still accepted, so existing configs keep working.

Use it

1. Call vitna_help to learn the checks.
2. Before a risky action, call vitna_preflight with the action text (e.g. "rm -rf /"). You get allowed | blocked | flagged.
3. Treat blocked / flagged (and engagement deny / hold) as a STOP: get human approval, then proceed. VITNA evaluates and records; it does not enforce.

vitna_help            what VITNA is + how to use it (no account)
vitna_preflight       dangerous shell/file/DB/network action check before you run it
consent_check         personal-data processing legality before you act
breach_classify       incident reportability (deadline + recipient)
dpia_threshold_check  must you run a DPIA first
ai_act_classify       EU AI Act risk tier before you build/ship

Trial vs claimed

Trial keys run the checks but are rate-limited (25/day, 200 lifetime), return label-only results, and keep no signed evidence. Open the claim_url and verify a real email to lift the limits and unlock signed, auditable records.

Honesty

VITNA is a cooperative guardrail, not a sandbox. vitna_preflight is heuristic pattern matching, so novel or obfuscated payloads can pass. It cannot block on its own; enforcement is yours.