HashtagWorld

SAFER is an open-source agent control plane that turns any AI agent running on LangChain, Google ADK, AWS Strands, raw Anthropic, raw OpenAI, OpenAI Agents SDK, AWS Bedrock, or CrewAI into one observable, governable, audit-ready system. The pain it solves: every team picks a different agent framework, ships an agent, and inherits three independent gaps. There's no shared view of what the agent did (each SDK ships its own callbacks and trace IDs), no shared way to enforce policy (every team rebuilds PII redaction and tool whitelists in code), and no shared report that survives an audit. SAFER is the missing layer. A two-line adapter normalises any framework into a 10-hook lifecycle contract that streams into a self-hosted FastAPI + SQLite WAL backend. Every event flows through three engines: a deterministic Gateway that enforces PII regexes and natural-language-compiled policies before tool calls, a Multi-Persona Judge (six personas, dynamically routed, single Opus 4.7 call per hot hook with > 80% prompt-cache hit rate), and a per-step Haiku 4.5 scorer that gates expensive Opus calls. Onboarding: an Inspector built as a Claude Managed Agent statically reviews the agent's source the moment it registers running grep + bash + AST scans inside a cloud sandbox, with a shared `safer-inspector-knowledge` memory store that gets smarter every scan. Pre-deploy: a Red-Team Squad (Strategist → Attacker → Analyst, also via Managed Agents) probes the live agent against OWASP LLM Top 10. Post-run: a deterministic Python aggregator + Sonnet 4.6 Quality + Thought-Chain Reconstructor produces a Session Report and a regulator-ready Compliance Pack (GDPR / SOC 2 / OWASP LLM) in PDF, HTML, or JSON. Apache 2.0, runs entirely in your VPC, never phones home. Ships in one `docker compose up`, or as `pip install safer-sdk` + `pip install safer-backend` (PyPI v0.1.1). 401 tests passing across the SDK + backend. Hi team, A quick note I'd like to flag with my submission for full transparency. I built this project together with a teammate on GitHub; both of our commits and pushes are visible on the repository. When I went to submit the form, I learned for the first time that team members must also be among the 500 selected participants. I wasn't aware of this rule until that moment, otherwise we would have handled the registration accordingly. Sharing this proactively, and happy to follow whatever process you advise. Thanks, Emine Hanedar