aifw

<div align="center"> <picture> <source media="(prefers-color-scheme: dark)" srcset="assets/securetron-logo.png"> <img src="assets/logo-light.svg" width="96" height="96" alt="AI-FW logo"> </picture> # AI-FW #### The AI Firewall & Governance Gateway **Every AI agent: identified, authorized, inspected, governed.** [Documentation](https://aifw.io/docs) · [Website](https://aifw.io) · [Get started](https://aifw.io/get-started) · [Enterprise](https://aifw.io/contact) [](https://aifw.io) [](https://aifw.io) [](https://aifw.io) [](https://aifw.io) [](https://aifw.io/docs) [](https://github.com/securetron-gh/aifw) </div> 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. - [x] **Fail-closed by design**: offline guardrails never delegate to third parties, no traffic while scanning is unhealthy - [x] **Agents are identities**: A2A protocol, agent registry, mTLS, Kerberos, OIDC SSO, SCIM provisioning - [x] **Cost savings without safety holes**: semantic-gated prompt compression and inspection-gated caching - [x] **Audit you can hand to a regulator**: metadata-only transaction log with encrypted export to your SIEM - [x] **Surfaces others cannot reach**: Claude inference hooks, MCP tool governance, M365 Copilot admin with agent sync - [x] **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](https://aifw.io/docs) and in this repository's `docs/` folder. Legend: ✅ native · 🧩 via partner or plugin · ❌ not supported | Category | Feature | Doc | Native | |---|---|---|---| | Security & guardrails | [Prompt &amp; response guardrails](docs/guides/01-prompt-response-guardrails.mdx) | [Read](https://aifw.io/docs/guides/prompt-response-guardrails) | ✅ | | Security & guardrails | [Semantic intent analysis](docs/guides/03-semantic-intent-analysis.mdx) | [Read](https://aifw.io/docs/guides/semantic-intent-analysis) | ✅ | | Security & guardrails | [Fail-open options (opt-in)](docs/guides/01-prompt-response-guardrails.mdx) | [Read](https://aifw.io/docs/guides/prompt-response-guardrails) | ✅ | | Security & guardrails | [Claude inference hooks](docs/guides/08-claude-inference-hooks.mdx) | [Read](https://aifw.io/docs/guides/claude-inference-hooks) | ✅ | | Identity & agents | [Identity &amp; access (IDAM)](docs/guides/04-identity-access.mdx) | [Read](https://aifw.io/docs/guides/identity-access) | ✅ | | Identity & agents | [Admin-issued API keys](docs/guides/11-agent-api-keys.mdx) | [Read](https://aifw.io/docs/guides/agent-api-keys) | ✅ | | Identity & agents | [A2A agent protocol](docs/api-reference/03-a2a-agent-protocol.mdx) | [Read](https://aifw.io/docs/api-reference/a2a-agent-protocol) | ✅ | | Identity & agents | [Agent self-enrollment](docs/tutorials/04-agent-self-enrollment.mdx) | [Read](https://aifw.io/docs/tutorials/agent-self-enrollment) | ✅ | | Cost | [Completion cache](docs/guides/06-reliability-caching.mdx) | [Read](https://aifw.io/docs/guides/reliability-caching) | ✅ | | Cost | [Prompt compression](docs/guides/07-prompt-compression.mdx) | [Read](https://aifw.io/docs/guides/prompt-compression) | ✅ | | Observability & audit | [Observability &amp; dashboard](docs/guides/09-observability-dashboard.mdx) | [Read](https://aifw.io/docs/guides/observability-dashboard) | ✅ | | Observability & audit | [Risk profiles &amp; auto-block](docs/guides/05-risk-profiles.mdx) | [Read](https://aifw.io/docs/guides/risk-profiles) | ✅ | | Observability & audit | [Audit logs &amp; export](docs/guides/10-audit-logs-export.mdx) | [Read](https://aifw.io/docs/guides/audit-logs-export) | ✅ | | Integrations | [M365 Copilot admin](docs/tutorials/03-m365-copilot-bridge.mdx) | [Read](https://aifw.io/docs/tutorials/m365-copilot-bridge) | ✅ | | Integrations | [Claude Code, Cursor &amp; MCP](docs/tutorials/02-claude-code-cursor-mcp.mdx) | [Read](https://aifw.io/docs/tutorials/claude-code-cursor-mcp) | ✅ | | Integrations | [Endpoint gateway](docs/api-reference/01-openai-compatible-api.mdx) | [Read](https://aifw.io/docs/api-reference/openai-compatible-api) | ✅ | | Core & reference | [Model routing &amp; registry](docs/guides/02-model-routing.mdx) | [Read](https://aifw.io/docs/guides/model-routing) | ✅ | | Core & reference | [Reliability](docs/guides/06-reliability-caching.mdx) | [Read](https://aifw.io/docs/guides/reliability-caching) | ✅ | | Core & reference | [Advanced parameters](docs/how-to/01-configure-models-keys.mdx) | [Read](https://aifw.io/docs/how-to/configure-models-keys) | ✅ | | Core & reference | [Prompt Analyzer](docs/admin/03-rules-manager.mdx) | [Read](https://aifw.io/docs/admin/rules-manager) | ✅ | | Core & reference | [Admin pages](docs/admin/01-settings.mdx) | [Read](https://aifw.io/docs/admin/settings) | ✅ | | Core & reference | [Supported models &amp; endpoints (200+)](docs/api-reference/04-openai-compatible-endpoints.mdx) | [Read](https://aifw.io/docs/api-reference/openai-compatible-endpoints) | ✅ |             > [!WARNING] > Model IDs change frequently. Always verify against each vendor's live model list before registering a model (see the [models reference](https://aifw.io/docs/api-reference/openai-compatible-endpoints)). ## ⚡ Quick start ### 1. Run the gateway and its database ```bash docker compose up -d ``` The [`docker-compose.yml`](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`. <details> <summary>Running the container without Docker Compose</summary> The image can be started on its own, but it does **not** provision a database, so you must supply a reachable PostgreSQL: ```bash 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`. </details> ### 2. Make your first request ```bash 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](https://aifw.io/docs/how-to/configure-models-keys)). ## 🔄 How it works ```mermaid 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](https://aifw.io/docs). ```mermaid 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"] G --> G1["settings, inventory, rules, risk, audit, dashboard, M365, users, keys, agents, CA, trust"] H --> H1["license & monthly token allowance"] ``` | Section | Topics | Files | |---|---|---| | Getting Started | What AI-FW is, quick start | `docs/getting-started/` | | Guides | Guardrails, routing, semantic analysis, identity, risk, reliability, compression, hooks, observability, audit, API keys | `docs/guides/` | | Tutorials | OpenAI SDK, Claude Code & MCP, M365 Copilot admin, agent self-enrollment | `docs/tutorials/` | | How-To | Models & keys, AI-assisted rules, Azure deployment | `docs/how-to/` | | API Reference | OpenAI-compatible, Anthropic Messages, A2A agent protocol | `docs/api-reference/` | | Admin Reference | Settings, Model Inventory, Rules Manager, Risk Profiles, Audit, M365, Users, Agents, CA, Agent Trust | `docs/admin/` | | Community Edition | [License & monthly token allowance](https://aifw.io/docs/community-edition/license) | `docs/community-edition/` | ## 🧩 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](https://aifw.io/contact) or [get started](https://aifw.io/get-started). ## 🔒 Security To report a security issue, contact us at [aifw.io/contact](https://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.