Open-source compliance monitoring for AI agents.
From Claude Code to Cowork — auto PII detection, risk scoring,
and GDPR / AI Act / SOC 2 mapping. Self-host or use our cloud.
from agentaudit import AgentAudit audit = AgentAudit(api_key="aa_live_xxxxx") event = audit.log( agent_id="booking-agent-v2", action="access_customer_record", data={"customer_email": "[email protected]", "fields": ["email", "phone"]}, reasoning="Customer requested booking modification", ) print(event.risk_level) # → "medium" print(event.pii_detected) # → True print(event.frameworks) # → {"gdpr": ["art_30"], "ai_act": ["art_14"]}
curl -X POST https://api.agentaudit.dev/v1/events \ -H "Authorization: Bearer aa_live_xxxxx" \ -H "Content-Type: application/json" \ -d '{ "agent_id": "booking-agent-v2", "action": "access_customer_record", "data": {"customer_email": "[email protected]"}, "reasoning": "Customer requested booking modification" }' # → {"risk_level": "medium", "pii_detected": true, "frameworks": {"gdpr": ["art_30"]}}
// .claude/settings.json — every action logged automatically { "hooks": { "PreToolUse": [ { "matcher": "", "hooks": [{ "type": "command", "command": "agentaudit-hook pre" }] } ], "PostToolUse": [ { "matcher": "", "hooks": [{ "type": "command", "command": "agentaudit-hook post" }] } ] } } // Zero tokens. Zero code changes. Every action captured.
AI agents are accessing customer data, running shell commands, querying production databases, and browsing confidential documents — right now, inside your infrastructure.
Anthropic's own audit tools don't capture Cowork activity. LangSmith tracks tokens, not actions. Zenity costs six figures and takes six months.
AgenticAudit fills the gap. Open source. Self-hostable. Audit-ready in 15 minutes.
Hooks fire on every action — the agent can't skip it. File reads, shell commands, API calls, connector access. All captured.
Auto PII detection — emails, IPs, API keys, credit cards. Risk scoring from low to critical. No configuration needed.
GDPR Art. 30, AI Act Art. 14, SOC 2 CC6.1 — mapped automatically. Your compliance officer gets reports, not spreadsheets.
| Time | Agent | Action | Risk | PII | Framework |
|---|---|---|---|---|---|
| 14:02:03 | claude-code | shell_commandpsql -h prod-db SELECT email FROM users | ⚠ HIGH | ✓ | Art. 30 |
| 14:03:44 | cowork | connector_accessgoogle_drive:read "Employee Comp Analysis.xlsx" | ⚠ HIGH | ✓ | Art. 15 |
| 14:05:12 | claude-code | file_read.env.production | ● CRIT | ✓ | Art. 30 |
| 14:05:30 | claude-code | shell_commandnpm test | ○ LOW | ✗ | — |
| 14:06:01 | cowork | connector_accessweb_search: "market analysis Q1 2026" | ○ LOW | ✗ | — |
Every file read, shell command, and code edit — logged automatically. Admins push config centrally. Developers can't opt out. One JSON config. Zero token overhead.
Every connector access, file operation, and web browse — audited via OpenTelemetry. Google Drive, Salesforce, DocuSign, Gmail — all captured. Point Cowork's OTLP endpoint to AgenticAudit. Zero setup for users.
Emails, IPs, API keys, credit cards, connection strings — flagged automatically. No configuration needed.
Rules-based engine: low, medium, high, critical. Based on action type, data sensitivity, and environment.
Four levels: minimal (GDPR-only), standard, full, paranoid (real-time blocking). Changes propagate instantly.
GDPR Art. 13/15/17/22/30, AI Act Art. 9/13/14, SOC 2 CC6.1/CC6.5/CC7.2 — mapped automatically to every event.
Append-only event store. Events can never be modified or deleted via API. Built for auditor trust.
Self-host it. Inspect every line. Your audit data never leaves your infrastructure unless you want it to.
Only logs events where PII is detected. Everything else is ignored. Lightweight, low noise.
All data access and security-sensitive operations. Routine actions skipped. Good signal-to-noise ratio.
Complete timeline of every action every agent takes. Nothing skipped. Full audit trail.
Real-time blocking via hooks. High-risk actions stopped before execution. Maximum protection.
| Free | Pro | Team | |
|---|---|---|---|
| Price | $0/mo | $49/mo | $199/mo |
| Events | 10K/mo | 100K/mo | 1M/mo |
| Retention | 30 days | 1 year | 2 years |
| Auto reports | — | ✓ | ✓ |
| Per-team policies | — | — | ✓ |
| RBAC / SSO | — | — | ✓ |
| SLA | — | — | 99.9% |
The open-source version is the full product, not a demo. Cloud adds managed hosting, retention, reports, and enterprise features.
| Agent | Support | Method | Enterprise Enforceable |
|---|---|---|---|
| Claude Code | ✅ Full (10/10) | Deterministic hooks | ✅ Policy settings |
| Cowork | ✅ Full (10/10) | Native OTLP | ✅ Org settings |
| LangChain | ✅ Full | Callback handler | — |
| CrewAI | ✅ Full | Event hook | — |
| Codex | 🟡 Partial (4/10) | Transcript parsing | ❌ Not yet |
| Cursor | 🟡 Partial (5/10) | Hooks + MCP | ❌ Unclear |
| Any agent | ✅ Full | REST API / Python SDK | — |
We're honest about what works. Claude Code and Cowork have 10/10 integration. Others are catching up.