AgentGuard CLI

Security testing for AI agents, before production. agentguard performs an offline static check or evaluates an agent through an AgentGuard control plane, prints an honest verdict, and exits non-zero when a deploy must be blocked or coverage is incomplete.

The judgement is deterministic (never an LLM deciding what's a vulnerability) and reproducible: the same configuration always yields the same verdict, explainable to the engineer it blocks at 2am.

Install

pip install agentguard-dev

The CLI has two lightweight dependencies (httpx and PyYAML). Offline static checks need no API key, network, database, or control plane.

First local check

agentguard init
agentguard scan --local

Static mode validates the manifest, declared policies, and accidental credentials. It does not execute a model, so behavioural scenarios are reported as SKIPPED and the command exits 40 (INCOMPLETE). To explicitly accept a static-only partial gate in CI:

agentguard scan --local --allow-incomplete-static --sarif agentguard.sarif

Behaviour simulation through a control plane

export AGENTGUARD_API_KEY=ag_your_key_here     # from your org's onboarding

agentguard scan \
  --api-url https://your-agentguard-host \
  --agent my-support-bot \
  --manifest manifest.json \
  --environment prod \
  --html report.html

Exit codes are the CI contract:

--html report.html writes a self-contained report (no external requests) with the verdict, evidence, and a per-finding remediation.

Compute a fingerprint locally

agentguard fingerprint --manifest manifest.json

The fingerprint identifies an exact agent configuration; a verdict is bound to it, so v2 never inherits v1's pass.

More

agentguard --help lists every command. Full docs and the control-plane setup live in the repository.