AI agent permission auditor
Check what an AI agent can do before you ship it. Paste its tool definitions, OpenAPI spec, MCP config, OAuth scopes or IAM policy, plus an optional manifest, and see every action it can take, where it can do more than the user it acts for, and how to fix it. Everything runs in your browser.
- 15 rules
- 8 formats
- 0 uploads
- PNG scope matrix
Prefer the terminal? npx @qlucent/agent-audit tools.json agent.yaml runs the same audit in CI. v0.1.0 · Suggest a rule
Runs in your browser. Nothing is uploaded.
Try a sample:
That doesn't make it safe: this checks what the spec declares, not what the agent's code does.
What the agent can do
Scope matrix
- allowed the role may do this
- exceeds the agent can, the role can't
- exceeds-minor a plain read beyond the role
- declared granted as the user's own records; the spec can't show the binding
Findings
FAQ
Do my specs get uploaded?
No. The audit runs in your browser, and the page's Content Security Policy blocks all network requests (connect-src 'none'), so nothing you paste or drop can leave the page. Nothing is stored, and descriptions, comments and other free text are never shown in the report.
Which files can I check?
OpenAI and Anthropic tool definitions, an MCP tools/list dump, OpenAPI 3 specs, OAuth scope lists, AWS IAM policies, MCP client configs (checked with the MCP config auditor's rules too) and an optional agent manifest. JSON, JSONC and YAML all work. Drop several files at once and they are audited as one agent.
What does the manifest add?
A small agent.yaml says who the agent acts for, what each of those roles may do, which actions need approval and what is out of scope. With it, the matrix compares the agent against each role and marks where the agent can do more than the person behind it. Without it you get an effects grid and approval rules fire one level lower.
How does it decide what an action does?
From names, parameter names, HTTP methods, MCP annotations and a catalogue of public OAuth scopes and AWS IAM actions. It is not an AI model, so it misses unfamiliar names: those are listed as unclassified for you to review, and you can declare effects in the manifest's actions block.
Can I run it in CI?
Yes. Run npx @qlucent/agent-audit with your spec files. It uses the same rules, exits non-zero on high or critical findings (--fail-on high by default), and can write SARIF for GitHub code scanning.