Skip to main content
Agentegrity evaluates an agent from inside its own decision loop. An adapter hooks into the framework you already use, turns each prompt, tool call, tool result and handoff into an event, and scores it through four layers: adversarial, cortical, governance and recovery. Every evaluation and every decision is appended to a hash-linked attestation chain that you can sign and verify later. The TypeScript packages report events and token usage without scoring them; the evaluation runs in the Python library and the coding-agent hook runtime. The library runs locally and blocks nothing unless you turn enforcement on. Guardrails and runtime monitors filter what goes in and out of an agent; Agentegrity measures the assembled agent (model, prompts, tools, memory and peers) and produces evidence of what it did and why it was allowed.

Quickstart

Instrument an agent in three lines and read its session report.

How it works

The four layers, the integrity properties and how a score is built.

Integrations

Eight Python adapters, six TypeScript packages and two coding-agent plugins.

Stream to Agentegrity Pro

Send sessions, verdicts and token usage to the console.

​
What you get from a session

  • Scores and verdicts. Each event gets a composite score and an action (allow, log, alert, block or escalate), with the reason recorded. See Scores and actions.
  • Decision provenance. Decision records are captured at each decision boundary before the action runs, and can be verified from the command line. See Attestation.
  • Token usage. The tokens a session spends, normalized across providers and totalled per model. See Token usage.
  • Multi-agent structure. Teams, crews, graphs and handoffs are declared as a topology that the chain commits to. See Multi-agent systems.

​
Supported frameworks

Claude Code and Codex are covered by hook plugins on a shared runtime. See Claude Code and Codex.

​
Open source

The library is Apache 2.0 and the specification is open. The source, specification and changelog are on GitHub.