This is a server whose users are agents, so its own tools are written on the assumption that a model might call them wrongly. Policies are created disabled so nothing can arm enforcement unattended. Quarantine is reversible and revocation is marked irreversible in the tool description itself. Tokens are short-lived by construction. Whether you need agent governance is a separate question — but the design here answers the obvious objection to putting governance behind an agent-callable API.
A governance control plane for other agents, reached over MCP. We connected and it listed 22 tools — identify itself as axiorank 1.0.1 — covering pre-execution scoring, policy, audit and incident response. Listing is public; every call needs a token.
- axiorank_score_tool_call — inspect a proposed call before executing it, returning a 0-100 risk score and a verdict of allow, deny or hold
- axiorank_verify_card — vet a remote agent or server's identity card before connecting to it
- Agent lifecycle: create with a short-lived bootstrap token, issue scoped tokens, quarantine reversibly, revoke permanently
- Outbound policies with tool patterns, risk thresholds and enforcement modes
- Audit search, incidents with the kill-chain behind them, ML threat assessments, and network-wide threat intel
OAuth2 through axiorank.com, or a bearer key beginning axr_live_. Scopes are per-area — policies:read and so on — and the tool list is readable before you have any of them.
Paste a URL, then authorize — add the endpoint to your client, then approve the OAuth consent screen