SidClaw
A policy and approval layer that intercepts AI agent tool calls, checks them against configured rules, and holds risky actions for human review before they execute, with integrations for MCP, LangChain, OpenAI Agents and Claude Agent SDK.
SidClaw sits between an AI agent and the tools it calls. Every tool call is evaluated against configured policies and resolves to one of three outcomes: allowed and executed immediately, held for a human reviewer who sees the full context and approves or denies it, or denied outright before the tool runs. Every decision is recorded as a trace.
It integrates with more than a dozen frameworks and platforms, including MCP, LangChain, OpenAI Agents, the Claude Agent SDK and Copilot Studio, so existing agents can add an approval layer without rewriting their tool-calling code. A local demo dashboard runs pre-loaded scenarios, such as a destructive shell command or a financial trade, without any signup or API key.
A self-hosted instance runs on Railway or Vercel with a Postgres-backed API and dashboard.