EU AI Act compliance
for MCP agents.

Pylae is a transparent proxy between your AI agents and MCP servers. It intercepts every tool call, enforces policies, and evaluates your infrastructure against 7 EU AI Act articles, generating the compliance evidence auditors will ask for.

Docker$docker run -p 8080:8080 ghcr.io/pylaelabs/pylae demo

Open source 7 EU AI Act articles 1,721 tests Self-hosted MIT / Apache 2.0

AI agents are in production. Governance isn't keeping up.

0
of organizations reported confirmed or suspected AI agent security incidents
0
of AI agents went live with full security and IT approval
0
of deployed agents operate without any security oversight or logging

Source: Gravitee State of AI Agent Security 2026 (N=919)

Enforcement starts August 2, 2026

And now the EU AI Act makes governance mandatory.

High-risk AI systems must meet Articles 9-15 by August 2, 2026. Non-compliance carries fines up to €15 million or 3 % of annual turnover. If your agents score credit, screen candidates, price insurance, or classify emergency calls, they fall under Annex III. Pylae evaluates your MCP infrastructure against all 7 articles and generates the compliance evidence auditors expect.

Article Requirement What Pylae evaluates
Art. 9 Risk Management System Active policies with risk-differentiated actions (block, escalate, rate_limit)
Art. 10 Data and Data Governance GDPR erasure capability, data retention configuration
Art. 11 Technical Documentation Agent Behavioral Contracts, policy versioning, system configuration
Art. 12 Record-keeping SHA-256 audit chain coverage, hash integrity, log retention period
Art. 13 Transparency and Provision of Information Decision source tracing, policy evaluation logging per action
Art. 14 Human Oversight Escalation policies, kill switch availability, dashboard authentication
Art. 15 Accuracy, Robustness, Cybersecurity Circuit breaker, description firewall, tool pinning, adaptive posture

Output: per-article compliance status, evidence list, gap analysis, numeric score, and overall readiness badge. Available via API and dashboard.

How Pylae works

1

Intercept

Pylae sits between your agents and their MCP servers. Every tool call passes through the proxy. Zero agent modification.

2

Evaluate

Policies define what's allowed. Block, escalate, rate-limit, or transform any action. YAML rules with glob matching.

3

Protect

8 security subsystems run on every request: firewall, tool pinning, circuit breaker, blast radius, adaptive posture, and more.

4

Comply

Every action is logged to a tamper-evident SHA-256 chain. Pylae evaluates your setup against 7 EU AI Act articles automatically.

AI Agent Claude, Cursor, etc. Pylae security · policies · contracts compliance · logs · alerts MCP Servers GitHub, Stripe, FS...

What a policy looks like

name: block-destructive-operations
action: block
priority: 100
match_rules:
  method: tools/call
  tool_name: "delete_*"   # glob pattern

Who uses Pylae

Works with any MCP-compatible agent. No agent modification needed.

Engineering teams

Drop-in proxy, no SDK, no agent modifications. Block delete_* and drop_* calls. Rate-limit per agent. Route to multiple MCP servers. YAML policies with hot-reload.

Compliance officers & DPOs

Automated EU AI Act readiness assessment across 7 articles. Compliance reports with evidence and gap analysis. GDPR erasure with PDF certificate. Insurability reports for cyber insurance.

CISOs & security teams

8 runtime security subsystems. Tamper-evident SHA-256 audit chain. Adaptive posture escalation. Incident replay with forensic export. Health score across 6 dimensions. Self-hosted. Your data never leaves your infrastructure.

Built for governance, not just monitoring

Compliance & Insurability Pro

EU AI Act readiness assessment across 7 articles with per-article scoring and gap analysis. Insurability reports (9 sections, SHA-256 sealed) for cyber insurance underwriting. GDPR erasure with PDF certificate. Forensic export.

8 Runtime Security Subsystems

Description firewall (30+ patterns). Tool pinning with hash tamper detection. PII sanitization. Blast radius scoring per session. Circuit breaker for agent loops. Adaptive posture (NORMAL→CRITICAL). Auto-remediation. Cross-server data flow control.

Agent Behavioral Contracts

Declarative YAML contracts defining what each agent can and cannot do. Forbidden tools, cost ceilings, escalation rules. Append-only versioning with side-by-side diff viewer and one-click restore.

Tamper-Evident Audit Trail

SHA-256 hash chain with Merkle block integrity. Every action cryptographically chained to the previous one. Chain verification, gap detection, and coverage percentage. CSV export.

Policy Engine

YAML policies with glob matching. Six action types: allow, block, escalate, rate-limit, log-only, transform. Hot-reload without restart. Append-only versioning with diff viewer.

Human-in-the-Loop

Escalate risky operations for human approval. Configurable timeout with secure token verification. Kill switch for all traffic, per-agent, or per-server. Works even if the DB is down.

Incident Replay & Forensics Pro

Full-screen timeline reconstruction of security events with context, chain verification, and policy evaluation trace. Natural language incident narratives. Forensic export with SHA-256 seal.

Cost Ceiling & Policy Simulation Pro

Per-agent budget limits (hourly/daily/monthly) with pre-ceiling alerts at 80%, 90%, 95%. Test policies against historical actions before deploying to see exactly what would change.

Real-Time Dashboard

10 pages: overview, actions, agents, policies, contracts, escalations, security, tool pins, compliance, settings. Health score across 6 dimensions. WebSocket live streaming. Actionable recommendations.

Multi-Server Proxy

Connect unlimited MCP servers through a single endpoint. Stdio and Streamable HTTP transports. Aggregated tool lists with transparent routing. CVE feed for known server vulnerabilities.

Notifications & Alerts Pro

In-dashboard notifications for security events. Webhook callbacks, Slack messages, and email alerts for escalations, security events, and policy violations.

Zero Config Install

Single Rust binary with embedded dashboard. No runtime dependencies, no containers required. Auto-registers agents, hot-reloads policies, defaults to allow. Self-hosted by design.

What no other MCP tool does

7 EU AI Act Articles

Automated compliance evaluation with evidence, gaps, and per-article scoring. No other MCP governance tool offers this today.

Insurability Reports

9-section governance posture report with SHA-256 cryptographic seal. Built for cyber insurance underwriting

Agent Contracts

Declarative YAML defining per-agent authorization boundaries. Policy ∩ contract = most restrictive wins. Versioned with diff viewer

Self-Hosted by Design

Your audit trails and compliance evidence never leave your infrastructure. Single binary, no cloud dependency. Data sovereignty built in

See it in action

The dashboard ships embedded in the binary. One process, one port.

🔒 localhost:8080
Pylae dashboard overview. Real-time activity chart, health score, and decision breakdown Pylae actions. Full audit log with decision source, policy trace, and security flags Pylae agents. Connected agents with session tracking, action counts, and status Pylae escalations. Pending human approvals with context, timeout, and approve/reject Pylae security. Health score across 6 dimensions, threat detection, adaptive posture Pylae compliance. EU AI Act readiness badge with 7-article evaluation and gap analysis Pylae policies. YAML rules with glob matching, version history, and simulation Pylae contracts. Agent behavioral contracts with YAML editor and version diff

Real-time overview with activity chart and decision breakdown.

Running in 60 seconds

Single binary, no runtime dependencies. The dashboard is built in. One process, one port, zero config to get started.

Simple, honest pricing

Self-hosted is always free. Runtime security included. No per-agent pricing, ever.

Free
€0
forever
  • Full MCP proxy with all 6 policy actions
  • 8 runtime security subsystems
  • Unlimited agents, servers & policies
  • Agent Behavioral Contracts (ABCs)
  • Kill switch & cost ceiling
  • Real-time dashboard (10 pages)
  • Human-in-the-loop escalations
  • 30-day action retention
Install Free

€99/mo to prove compliance. EU AI Act fines reach €15 million or 3 % of annual turnover.
Need multi-user or managed hosting? Contact us