Open-Source AST Tool Blocks Dangerous Agent Commands in Under 40 Microseconds

Developers frustrated with cloud-based safety guardrails for local AI agents have released an open-source Python library called btp-guard, available on PyPI. The tool uses an in-process Abstract Syntax Tree (AST) evaluator to inspect Python, SQL, and Bash code blocks directly in memory before execution, eliminating the need for remote moderation APIs. Unlike cloud-based solutions such as AWS Bedrock or OpenAI Moderation, which add 1,500–2,500 milliseconds of latency per call, btp-guard intercepts dangerous commands in under 40 microseconds with no OS system calls. The library is designed to block high-risk operations including catastrophic filesystem deletions, database purge commands, Python sandbox escapes, and credential exfiltration attempts. It targets developers running local large language models through tools like Ollama, vLLM, or LM Studio who require both speed and air-gapped privacy guarantees.