auditagent-mcp

An MCP server exposing your AuditAgent workspace as five tools: get_recent_actions, get_pending_approvals, decide_approval, draft_questionnaire_answers, get_compliance_summary. Works from Claude, ChatGPT, Grok, or any other MCP client — two different ways to connect, depending on which:

get_pending_approvals returns a ready-made summary sentence per item (e.g. "ops-agent wants to run bulk_delete_records (external) with {'count': 500} — requested ..., currently pending") alongside the flattened fields it's built from, so a conversation can relay it directly instead of narrating raw JSON.

decide_approval lets you approve or reject one of those from inside the conversation, as an alternative to clicking through the dashboard — but only with a reviewer key: create one in Settings → API keys with "Can approve/reject" enabled. An ordinary agent-tracking key is refused on purpose, since an agent must never be able to decide its own pending request.

Option A — hosted (Claude.ai web/mobile, ChatGPT, Grok)

These are hosted chat products with no local machine to run a server on, so they can only reach a server over the network, not installed via pip. This server is already mounted at /mcp on the deployed backend (see backend/app/main.py) — add <your backend's base URL>/mcp as a custom connector, e.g. this deployment's Hugging Face Space: https://awais1290-auditagent.hf.space/mcp:

• Claude.ai: Settings → Connectors → Add custom connector → paste the URL, and use your AuditAgent API key (Settings → API keys in the dashboard) as the bearer token when prompted.
• ChatGPT: Settings → Connectors → Developer mode (Plus/Pro and up) → add a custom connector with the same URL and bearer token.
• Grok: grok.com/connectors → New Connector → Custom → same URL and bearer token.

Each connection authenticates as your own workspace — nothing is shared between different users of the hosted server.

Option B — local (Claude Desktop, Claude Code)

These run as a local process on your own machine, so they can spawn this server directly instead of connecting to the hosted one.

pip install auditagent-mcp

Get a workspace API key from Settings → API keys in the AuditAgent dashboard, then set:

export AUDITAGENT_API_KEY=al_live_...
export AUDITAGENT_BASE_URL=https://awais1290-auditagent.hf.space   # optional, this is the default

Add to your MCP config (Claude Desktop: claude_desktop_config.json; Claude Code: .mcp.json or claude mcp add):

{
  "mcpServers": {
    "auditagent": {
      "command": "auditagent-mcp",
      "env": {
        "AUDITAGENT_API_KEY": "al_live_..."
      }
    }
  }
}

Using it

Once connected (either option), ask things like:

"What agent actions happened in the last hour?" "Are there any approvals waiting on me?" "Draft an answer to: do you log all AI agent actions taken on customer data?" "What's our current compliance posture?"

Manual setup: getting listed for organic discovery

Two separate, unrelated listings — do either or both:

• The official MCP Registry (registry.modelcontextprotocol.io, a community-run metadata index, not an Anthropic product): install the mcp-publisher CLI, run mcp-publisher init to generate a server.json, mcp-publisher login github, then mcp-publisher publish. This is what makes any MCP-aware client or aggregator find this server by name.
• Anthropic's own Connectors Directory (the curated list inside claude.ai/Claude Desktop/Claude Code's own UI): a separate submission through https://claude.ai/admin-settings/directory/submissions/new (remote servers like this one; needs a Team/Enterprise claude.ai org) — requires a documentation URL, a privacy policy URL, and OAuth for any authenticated action (this server currently uses a static bearer token instead, which the portal may or may not accept as-is; check its current requirements before submitting).

Both are manual, one-time submissions tied to your own identity/ownership — not something this repo automates.