MCP Audit
MCP server inspection, permission review, and connector risk scoring.
About
Inspects an MCP server endpoint, enumerates declared tools and their schemas, reviews permission scopes, looks for broad-or-unscoped tool definitions, flags credential requests, and emits a risk score and detailed finding report. Read-only: does not modify the MCP server.
Native install
For environments with file or shell access (Codex, Claude Code, Gemini CLI, Cursor, local CLI). v0.1: the apai CLI is a scaffold; real install behavior lands in Phase 1.
The apai CLI is a TypeScript scaffold at v0.1. See /honest-status for the full shipped vs stubbed list.
Prompt install
Hosted chat apps (ChatGPT, Claude, Gemini, Grok) cannot silently install software. Paste the snippet below into the chat to load the package protocol for the current conversation only.
Paste this into any LLM that does not have shell access to your machine - chatgpt.com, claude.ai, gemini.google.com, grok.com - and the model will load the Capability Passport and apply the package protocol. If you are using Claude Code, Codex, Cursor, Gemini CLI, or another tool-enabled agent, use apai install instead - it produces a durable install receipt.
Install https://apai.run/packages/mcp-audit for this chat. Load the capability passport. Given an MCP server URL, return a structured permission and risk audit. Do not authenticate against the server; describe the endpoint contract only.
Install card source: https://apai.run/packages/mcp-audit/llms.txt
Capability Passport
Reads
- ·Publicly readable MCP server tool listings
Writes
- ·Local audit report file
Accesses
- ·Network endpoints declared by the user
Spends
- ·No spend
Exposes
- ·Aggregated risk-finding summary if shared via cloud sync
Approvals required
- ·Approval required for authenticated audit (when added)
Rollback
Delete generated audit report.
Scanner
heuristic-stub-v0 means the v0.1 scanner (13 suspicious-pattern regex checks + Unicode format-character detection) found nothing. It does NOT mean the package is certified safe. Real prompt-injection, OAuth-scope, and dependency scanning lands in Phase 5. See honest status.