securetron-gh/aifw

AI Agent Identity and Access Management and Inference Firewall

What it solves

AI-FW addresses the need for security, governance, and auditability when AI agents talk to language models. Without it, organizations have little visibility into what prompts are sent, what responses come back, or whether policies are enforced. AI-FW places a gate in the middle so every prompt and response can be checked, governed, and logged, and it fails closed so no traffic flows when inspection is unhealthy. It also adds provable agent identity and audit trails that can be shared with regulators.

How it works

The gateway sits between clients (agents, SDKs, tools) and LLM backends. All traffic goes through three steps: inspect (scan prompts, mask PII, score semantic intent), govern (apply rules, risk profiles, and policies to decide what is allowed), and route (send to the correct model backend using a registry and resilience logic). Responses come back and an audit step records metadata-only logs of what happened. The gateway supports OpenAI-compatible and Anthropic endpoints, and integrates with Claude inference hooks, Cursor Agent Hooks, MCP tool governance, and Microsoft 365 Copilot admin. You start it with Docker Compose, which runs the gateway and a PostgreSQL database, then point any compatible client at the gateway's URL.

Who it's for

Organizations that run AI agents or copilots and need enterprise-grade security, compliance, and cost control. This includes teams using Claude Code, Cursor, MCP tools, Microsoft 365 Copilot, or any OpenAI-compatible application. It suits self-hosting users who want no per-token fees and full control, and those who must meet ISO 27001 or SOC2 requirements.

Highlights

  • Fail-closed design: no traffic flows while inspection is unhealthy; offline guardrails never delegate to third parties.
  • Agents are treated as identities with A2A protocol, agent registry, mTLS, Kerberos, OIDC SSO, and SCIM provisioning.
  • Cost savings via semantic-gated prompt compression and inspection-gated completion caching.
  • Regulator-ready audit: metadata-only transaction logs with encrypted export to SIEM via syslog over TLS.
  • Surfaces deep integrations: Claude inference hooks, Cursor Agent Hooks, MCP tool governance, M365 Copilot admin with agent sync.
  • Optional endpoint gateway for images, audio, files, fine-tuning, and batches under the same inspection pipeline.
  • Supports 200+ models and endpoints from OpenAI, Anthropic, Google, Meta, xAI, DeepSeek, Mistral, Groq, Azure, AWS, Alibaba, Ollama, and more.
  • Self-hosted with no per-token fees; enterprise features include role-based access control, OIDC SSO, SCIM, and encrypted audit export.

Related

  • Project
    menkesu/awesome-pm-skills

    Awesome PM Skills 2.0 is a collection of 25 product-management workflows distilled from 340 episodes of Lenny's Podcast. Each skill takes a real work artifact like a pricing page or PRD and guides an AI assistant to produce a concrete deliverable such as a pricing plan or decision memo, with expert frameworks, grading, and verified quotes.

  • Project
    SponsioLabs/Sponsio

    Sponsio is an open‑source Python/TypeScript library that adds deterministic, model‑free runtime checks to LLM‑powered agents. By compiling LTL‑style contracts into a fast monitor, it blocks unsafe tool calls (e.g., destructive SQL, shell commands) in microseconds, integrates with major agent frameworks, provides ready‑made contract bundles, and offers an optional hosted console for rule review.

  • Project
    sjtu-sai-agents/ML-Master

    ML-Master is an open-source AI agent that automates machine learning engineering tasks by integrating exploration and reasoning with an adaptive memory system. It ranks #1 on OpenAI's MLE-Bench, and version 2.0 adds a hierarchical cognitive caching architecture for ultra-long-horizon autonomy.

  • Project
    THUSI-Lab/GameScaling

    GameScaling is a set of vision-language models that were continued pre-training on 1,600 hours of keyboard-and-mouse gameplay so they can play computer games in real time. Each model takes the current screen (plus a short history) and outputs the next 200 ms of input as six 33 ms key steps plus one mouse move and scroll; the repo provides the inference client, system prompt template, examples, and tests for serving the released checkpoints.