AgentKey

AgentKey · Overview

What is AgentKey?

AgentKey is an authorization and evidence layer for AI agents. It sits between an autonomous agent and the external tools and APIs the agent calls, evaluates every action against a policy before the action runs, records each decision to a tamper-evident audit trail, and produces verifiable proof of what the agent did.

By AgentKey · Last updated 2026-09-01

The problem

Every team shipping an AI agent eventually hits the same wall. The agent can call tools, send email, move money, and modify data, but you cannot reliably see what it did, you cannot stop it when it goes off script, and you cannot prove to an auditor or a customer what happened.

The agent itself is not the risk surface. The actions it takes against real systems are. AgentKey is built to control that surface.

How it works

AgentKey is a thin control plane you add to any existing agent. It is not a framework, not an agent runtime, and not a model. Every tool call flows through one authorization check, is recorded to an append-only evidence chain, and can be verified later without trusting AgentKey.

  • AI Agent issues a tool call (charge a card, send email, query a database).
  • AgentKey evaluates the action against the agent's permission policy and risk signals.
  • The decision is Allow, Block, or Require human approval.
  • The action reaches the external tool or API only if allowed.
  • Every decision and execution is written to a hash-chained evidence session.
  • Anyone with the session token can recompute the hashes and confirm the record is intact.

What it provides

Four capabilities work together:

  • Control: decide which actions each agent may take, down to the parameter. Block a charge over $100. Allow refunds under $500 with approval. Restrict email sends to your own domain.
  • Observe: every decision and execution is written to a hash-chained evidence ledger with a replayable session timeline.
  • Prove: the evidence chain is tamper-evident by construction. Verification recomputes hashes from raw events without trusting AgentKey.
  • Trust the supply chain: a tool trust registry scores and flags every MCP server, API, and plugin the agent connects to.

An example policy

Policies are plain JSON matching a resource and action, with optional parameter conditions. No permission configured means the default is deny.

json
{
  "resource": "stripe.charge",
  "action": "invoke",
  "decision": "allow",
  "conditions": [
    { "field": "amount", "operator": "lte", "value": 100, "on_fail": "deny" }
  ]
}

Example use cases

  • A support agent that reads and sends email, where every send is either allowed by policy or paused for a human.
  • A billing agent that can issue refunds under a threshold automatically and requires approval above it.
  • An operations agent connected to a production database, restricted to read actions and a specific table scope.
  • A research agent that ingests web pages and email, scanned for prompt injection before the content reaches the model.

Where AgentKey fits

AgentKey is the product this site documents. The rest of the Learn pages explain each concept independently, then show where AgentKey implements it. Start with the docs for installation and the SDKs.

Frequently asked questions

What is AgentKey?

AgentKey is an authorization and evidence layer for AI agents. It evaluates agent actions against a policy before they reach external tools or APIs, records every decision to a tamper-evident audit trail, and lets anyone verify the record is intact.

Is AgentKey an agent framework or a model?

No. AgentKey is a thin control plane you add to an existing agent. It does not run your agent or replace your model; it controls and records what the agent does.

How do I prove what an agent did?

Every authorization and execution is written as a hash-chained evidence event. To verify a session, you recompute the hashes from the raw events and compare them to the stored root. If they match, the record is intact, without trusting AgentKey.

Does AgentKey block actions or just observe them?

Both. In observe mode it records and surfaces actions without blocking. In enforce mode it blocks denied calls and escalates risky decisions to a human approval.

Related topics

AI agent securityAI agent authorizationDocumentationAll topics