Getting started

Go to the app and sign in with your work email (magic link, no password to set). On first login you'll be asked to name a workspace. One workspace per team or product is typical.

In the dashboard, open Settings → API Keys. Click Create key, give it a name that says where it'll live (e.g. prod-checkout-agent), and copy the value shown. It's only shown once, so store it as an environment variable (AUDITAGENT_API_KEY) wherever your agent runs.

One decorator is the whole integration for logging:

pip install AudAgent

from audagent import AuditAgent

audit = AuditAgent(api_key="AUDITAGENT_API_KEY", agent_name="billing-bot")

@audit.track(action_type="external", action_name="send_refund")
def send_refund(customer_id: str, amount_cents: int):
    ...  # your existing function, unchanged

audit.track wraps your function without changing what it does or returns; it just reports that the call happened. Inputs and outputs are redacted for common PII, and a second automated pass catches secrets or API keys, before the event is ever stored.

Most actions don't need a human in the loop. Some, like refunds over a threshold or anything destructive, should. That's what Policy configures:

rules:
  - match:
      action_name: send_refund
      inputs.amount_cents: { gt: 10000 }
    require_approval: true

Once a rule matches, audit.track blocks (it polls, it doesn't spin) until someone approves or rejects, then your function runs, or doesn't, and the result is what your code sees.

Without Slack, pending approvals still work: they show up on the Approvals page and, with a fallback email set in Settings, as an emailed link. Most teams still connect Slack from Settings so anyone in the right channel can approve with one click.

Under Questionnaires, upload a customer's security or compliance questionnaire (PDF, DOCX, or plain text). AuditAgent parses out individual questions, finds your actual logged events that are relevant evidence for each one, and drafts a cited answer. You review and edit every answer before exporting; nothing goes to a customer without a human reading it first.

The SOC 2 page maps common Trust Services Criteria to what AuditAgent is logging for you. It's a starting point for your own audit prep, not a certification. Talk to an auditor before making that claim externally.

AuditAgent also runs as an MCP server, so you can ask an AI assistant things like “anything waiting on me?” or “what did the billing agent do last night?” directly. It reads the same record as the dashboard, and it can only read. Two ways to connect, depending on which assistant:

Claude.ai, ChatGPT, or Grok (hosted, no install): these run in the browser with no local machine to install anything on, so they connect to AuditAgent's own hosted MCP endpoint instead. Add a custom connector pointing at your deployed backend's /mcp path, using your AuditAgent API key as the bearer token when prompted.

Claude Desktop or Claude Code (local): these run as a process on your own machine, so they can run the server directly:

pip install auditagent-mcp

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