AI-FW

The AI Firewall & Governance Gateway

Every AI agent: identified, authorized, inspected, governed.

Documentation · Website · Get started · Enterprise

AI-FW is the AI firewall and governance gateway between your AI agents and the models they call. Every prompt and response is inspected, governed, and routed, with provable agent identity and regulator-ready audit, all in one process that runs in your own environment.

• Fail-closed by design: offline guardrails never delegate to third parties, no traffic while scanning is unhealthy
• Agents are identities: A2A protocol, agent registry, mTLS, Kerberos, OIDC SSO, SCIM provisioning
• Cost savings without safety holes: semantic-gated prompt compression and inspection-gated caching
• Audit you can hand to a regulator: metadata-only transaction log with encrypted export to your SIEM
• Surfaces others cannot reach: Claude inference hooks, Cursor Agent Hooks, MCP tool governance, M365 Copilot admin with agent sync
• Opt-in endpoint gateway: images, audio, files, fine-tuning, and batches under the same inspection pipeline

🛡️ Features

Every feature is documented on aifw.io/docs and in this repository's docs/ folder. Product updates and practical guidance are published on AI-FW Insights, with an RSS feed.

Legend: ✅ native · 🧩 via partner or plugin · ❌ not supported

[!WARNING]
Model IDs change frequently. Always verify against each vendor's live model list before registering a model (see the models reference).

⚡ Quick start

1. Run the gateway and its database

docker compose up -d

The docker-compose.yml in this repository starts the gateway alongside PostgreSQL: both share a network, the database runs on a named volume so data survives restarts, and the gateway waits for it to report healthy before starting. One command, no manual setup.

The default login is superadmin / happy.

[!NOTE]
The gateway listens on 80 (HTTP) and 443 (HTTPS). Pin a published version with AIFW_TAG=1.1.245 docker compose up -d.

The image can be started on its own, but it does not provision a database, so you must supply a reachable PostgreSQL:

docker run -d --name aifw-gateway -p 8080:80 \
  -e "ConnectionStrings__PostgreSQL=Host=your-postgres;Port=5432;Database=aifw;Username=...;Password=..." \
  ghcr.io/securetron-gh/aifw:latest

[!IMPORTANT]
Without a reachable PostgreSQL the gateway still starts, but it falls back to in-memory stores and all data is lost on restart. The built-in default connection string points at localhost, which inside the container is the container itself, so a sibling database container must be addressed by its name on a shared Docker network.

The same image is published to Docker Hub as securetron/aifw.

2. Make your first request

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"gpt-4o","messages":[{"role":"user","content":"Hello!"}]}'

Open the gateway at http://localhost:8080 and watch the transaction appear in the dashboard.

[!TIP]
To route traffic through the gateway, register these backend URLs and models in the AI-FW Model Inventory (see Configure models and keys).

🔄 How it works

flowchart LR
  A["Client / Agent"] --> B["AI-FW Gateway"]
  B --> C["Inspect<br/>prompt scan, PII mask"]
  C --> D["Govern<br/>rules, semantic tier"]
  D --> E["Route<br/>registry, resilience"]
  E --> F["LLM Backend"]
  F --> G["Audit<br/>metadata-only log"]
  G --> A

1. Connect - point any OpenAI-compatible or Anthropic client at the gateway (SDKs, Claude Code, Cursor, MCP tools, or plain HTTP)
2. Inspect - every prompt and response is scanned, PII-masked, and semantically scored
3. Govern - routing rules, risk profiles, and policy decide what flows and what is blocked
4. Audit - metadata-only logs and events record exactly what happened

📚 Documentation

This repository holds the AI-FW knowledge base. The rendered version lives at aifw.io/docs.

flowchart TD
  A["docs/"] --> B["getting-started/"]
  A --> C["guides/"]
  A --> D["tutorials/"]
  A --> E["how-to/"]
  A --> F["api-reference/"]
  A --> G["admin/"]
  A --> H["community-edition/"]
  B --> B1["what is AI-FW, quick start"]
  C --> C1["guardrails, routing, semantic, identity, risk, reliability, compression, hooks, observability, audit, API keys"]
  D --> D1["OpenAI SDK, Claude Code & MCP, M365, agent enrollment"]
  E --> E1["models & keys, AI-assisted rules, Azure"]
  F --> F1["OpenAI-compatible, Anthropic Messages, A2A, endpoints, current model catalog"]
  G --> G1["settings, inventory, rules, risk, audit, dashboard, M365, users, keys, agents, CA, trust, compliance"]
  H --> H1["license & monthly token allowance"]

🧩 Enterprise

• ✅ ISO 27001 certified and SOC2 compliant
• ✅ Fail-closed by design - no traffic while scanning is unhealthy
• ✅ App-only Microsoft 365 Copilot integration with agent sync
• ✅ Role-based access control, OIDC SSO, and SCIM provisioning
• ✅ Encrypted audit export to your SIEM (syslog over TLS)
• ✅ Self-hosted in your environment with no per-token fees

Book a demo or get started.

🔒 Security

To report a security issue, contact us at aifw.io/contact.

🤝 Contributing

PRs are welcome.

We would also like to thank the Securetron.net team for their continued support in helping the community.