Introduction

AxioRank is the Zero-Trust control plane for AI agents, in both directions. Outbound, every action an agent takes (a tool call, an outbound payment card, an MCP or A2A message) passes through AxioRank, where it is inspected, scored, and then allowed or denied, with any secrets in the payload redacted before it's logged. Inbound, the agents reaching into the surfaces you operate (a website, an HTTP API, an MCP server, an A2A agent, or a webhook) are verified and gated the same way. And discovery finds the AI usage that never went through AxioRank at all. Identity in, decision out.

The Zero-Trust model

Agents are fast, capable, and easily manipulated. A prompt-injected agent will exfiltrate a secret or fire a destructive command as readily as it completes the task you asked for. AxioRank treats every agent action as untrusted by default, no matter which agent makes it or how many times it has been allowed before:

1. Intercept: the agent's request is routed through the gateway instead of hitting the tool directly, whether through an SDK wrapper, an MCP shim, or the drop-in AI Gateway proxy.
2. Inspect: the payload is scanned for secrets, PII, destructive operations, and injection attempts.
3. Decide: a risk score and the active policy determine the verdict: allow, deny, or hold for human approval.
4. Record: every decision is logged with the signals that produced it, so the trail is auditable after the fact.

Core capabilities

• Content inspection: payload-aware detection of secrets, PII, destructive actions, and prompt injection, with write-time redaction. See Content-inspection engine.
• Protocol adapters: interoperability with the agent ecosystem (MCP, A2A) and outbound controls such as card preflight. See Protocol adapters.
• AI Gateway: a drop-in, OpenAI-compatible proxy that governs every model call across OpenAI, Anthropic, Bedrock, Vertex, and Gemini by changing one base URL. See AI Gateway.
• Risk scoring & policies: signals roll up into a score; policies turn the score into an enforceable verdict per agent, tool, and request attribute.
• Spend governance: attribute cost to every governed call, chart it by agent and model, and cap it with workspace budgets. See Spend governance.
• Detection intelligence: post-hoc ML assessment, taint provenance, and kill-chain correlation behind the hot path. See Detection intelligence.
• Human-in-the-loop approvals: hold risky calls for one-click email or Slack approval, with dual control and escalation. See Approvals.
• Inbound surfaces: verify the agents reaching into a website, HTTP API, MCP server, A2A agent, or webhook you operate. See Inbound surfaces.
• Shadow AI discovery: find ungoverned AI usage from egress logs, MCP scans, identity events, and cloud audit trails. See Shadow AI discovery.
• Assurance: one health view for integrations, the ML lane, data residency, and the async job queue. See Assurance center.
• Provable audit: the audit log is itself a hosted RFC 6962 Merkle transparency log with hourly signed tree heads, a per-action receipt you verify offline, and independent witnesses. See Audit integrity and Verify our log.
• Governance Copilot: ask AxioRank about your own agents, policies, incidents, and spend in plain language, grounded and read-only. See AI Governance Copilot.

Next steps

• Getting started: wire an agent through the gateway.
• Framework integrations: guard tools in the Anthropic SDK, Vercel AI SDK, LangChain, and more.
• Policies: turn risk into an enforceable verdict.
• Inbound surfaces: verify the agents that reach into what you operate.
• Content-inspection engine: how payloads are scored.