Built for teams that take AI governance seriously. Here's how each feature works and why it matters.
When you run a multi-agent system, you don't just have one AI making decisions — you have a whole team. A CrewAI crew might have a researcher, a writer, and an editor. A LangChain chain might call three different tools in sequence. If something goes wrong, you need to know which agent did what at which step.
AgentAudit gives every multi-agent workflow a trace ID. Every action gets logged with a parent span ID linking it to the step that triggered it. The result is a complete tree you can visualize, query, and export.
Most compliance tools work like a security camera: they record the violation after it happens. AgentAudit works like a security guard: it stops the violation before it reaches the user.
When an agent generates output, you send it through the guardrail endpoint. AgentAudit checks it against your compliance rules — PII detection, forbidden keywords, regex patterns, sentiment thresholds, custom validators. If a violation is found, the output is blocked and you get a detailed report of what was caught and why.
Blocking a violation is only half the battle. You also need to know it happened, who was affected, and how severe it was. AgentAudit sends real-time alerts on every violation with full context.
Alerts include the triggering action, the compliance rules that matched, the severity level (warning vs. critical), and the raw input/output that caused the flag. You can route these to Slack, email, or any webhook endpoint. Alerts are also stored in PostgreSQL for historical analysis and reporting.
When the auditor shows up, you need proof. AgentAudit generates audit-ready reports in CSV or JSON format for any date range, any agent, or any compliance rule type.
Reports include every log entry with its full metadata, compliance flags, trace IDs, and timestamps. You can filter by agent, action type, severity, date range, or custom metadata fields. Export in seconds, not hours.
Compliance tools handle sensitive data, so security isn't a feature — it's the foundation. AgentAudit is built on PostgreSQL with row-level organization isolation, meaning your data never mixes with another tenant's.
API keys are hashed with bcrypt and a unique salt. JWT tokens have configurable expiration. Rate limiting protects auth and audit endpoints from brute force. Custom JavaScript validators run in sandboxed contexts with 100ms timeouts to prevent ReDoS and infinite loops.
The Model Context Protocol is emerging as a standard for how AI agents exchange structured data. AgentAudit's REST API accepts the same structured JSON payloads that MCP-based agents already emit, so you can route agent events to AgentAudit without reformatting them.
Map your MCP tool-call results or context updates to audit log fields (action, prompt, response, metadata), and every event flows through the full compliance and tracing pipeline — no framework lock-in, no special adapter required.