Hi HN! I’m John, building Authensor, and I just put out SafeClaw.
SafeClaw is a local policy gate for AI agents. It hooks into tool calls (file writes, shell exec, network requests) and evaluates each one against a configurable policy before the action runs. Risky operations pause and wait for human approval via CLI or dashboard. If the control plane is unreachable, everything non-read is denied. Zero npm dependencies beyond the Agent SDK, runs entirely on your machine, your API key never leaves localhost.
Install:
npx @authensor/safeclaw
Opens a local dashboard + setup wizard (Node 20+). Works with Claude + OpenAI.
Notes: open-source MIT client, sub-ms local eval, dry-run mode, tamper-evident hash-chained audit log, 446 tests.
Happy to answer questions about threat model + design tradeoffs.
Hi HN! I’m John, building Authensor, and I just put out SafeClaw.
SafeClaw is a local policy gate for AI agents. It hooks into tool calls (file writes, shell exec, network requests) and evaluates each one against a configurable policy before the action runs. Risky operations pause and wait for human approval via CLI or dashboard. If the control plane is unreachable, everything non-read is denied. Zero npm dependencies beyond the Agent SDK, runs entirely on your machine, your API key never leaves localhost.
Install: npx @authensor/safeclaw Opens a local dashboard + setup wizard (Node 20+). Works with Claude + OpenAI.
Notes: open-source MIT client, sub-ms local eval, dry-run mode, tamper-evident hash-chained audit log, 446 tests. Happy to answer questions about threat model + design tradeoffs.