A local advisory classifier for prompt injection, excessive agency, destructive commands, supply-chain signals, and more. Your integration decides whether a verdict blocks execution.
Prompt injection, excessive agency, supply-chain attacks, error leakage, reality-frame override, and more. Each boundary is a focused detector.
No API calls, no telemetry, no data leaves your machine. The classifier is a Python module you import.
Call Agent Trust before execution, then enforce its receipt in your own non-bypassable policy boundary.
Read the code, run it yourself, contribute boundaries. Part of the Ouroboros project — an AI agent that builds itself.
A shell command, an API call, a file write — anything the agent wants to do.
Deterministic boundaries scan the action for known attack patterns, excessive permissions, and information leakage.
The library returns a sanitized recommendation. It does not execute or block the action itself.
Paste a prompt to try an illustrative browser-only rule set. The Python package is the canonical implementation, and this demo does not enforce anything.
We ran Agent Trust against Anthropic Skills mcp-builder SKILL.md — a real MCP server skill definition used in production.
Verdict: ALLOW — the document is clean. Four false positives on build-command keywords (run, publish, http, package, coverage) — the gate over-triggers on domain vocabulary. This is useful feedback for both the skill author and for us.
We offer free reviews for open-source agent projects. Send us your SKILL.md, CLAUDE.md, or agent prompt — we'll run the gate and share the results.
Install the Python package, call it at your pre-action boundary, and keep enforcement in code you control.
Read the quick start →