ButterClaw: Self-Hosted Runtime Security for AI Agents, No Cloud
ButterClaw enforces AI agent security locally with regex signatures, a local LLM verdict pipeline, and SIGKILL/credential shredding — no cloud, no telemetry.
ButterClaw is a self-hosted security layer for autonomous AI agents that intervenes in real time rather than just logging activity. It sits inside the agent's execution path — before the LLM call, after it, and before each tool invocation — using a local Ollama model for reasoning, so no data ever leaves the host machine.
The pipeline starts with a millisecond-fast regex-based "Arsenal" of seven attack signatures — covering reverse shells, credential exfiltration, prompt injection, and cloud metadata probes — that can kill a process before the LLM is even invoked. Suspicious activity that passes this gate goes through a two-stage verdict: a "Guardian Brain" evaluates behavioral drift using recent tool-call history, and a separate zero-temperature "Auditor" independently checks that verdict before any kinetic action, such as SIGKILL or credential revocation and shredding, is taken.
Unlike observability tools like LangSmith/LangFuse, which only log agent traces, or cloud-based products like Halo, ButterClaw enforces policy locally with a deterministic engine, six alert channels, and a reproducible 25/25 live-fire test suite. The Apache 2.0 licensed project needs only Docker and Ollama to run, and is seeking MCP-focused co-maintainers as part of its Agentic AI Foundation application.
This synthesis was produced from its source by AI; there is no human editor or manual review step. How we work