SidClaw is an approval and accountability layer for agentic AI, and its core process can be summarized as identity → policy → approval → tracking. It intercepts the moment an AI agent calls a tool, compares the call against preset policies, and reserves risky operations for human review before deciding whether to execute them. The tool emphasizes that it does not change users' existing agent logic, but instead wraps existing tool calls, thereby achieving governance at the tool-call layer.
In terms of main capabilities, SidClaw provides four primitives to govern every operation of an agent: allow means the operation executes immediately and is tracked; approval_required means a human reviews a context card and then approves or rejects it, and it is tracked; deny means it is directly blocked before execution, without accessing data, and it is likewise tracked. As an MCP governance proxy, it can wrap any MCP server and is compatible with MCP clients such as Claude Desktop, Cursor, VS Code, and GitHub Copilot. Platform features include a policy engine (allow / approval_required / deny, with priority ordering and classification hierarchy), manual approval workflows (context cards showing the agent's reasoning process, risk classification, and separation of duties), audit trails (correlated tracking with an integrity hash chain, tamper-resistant), and SIEM export (JSON and CSV, continuous webhook push). Capabilities for platform teams also include RBAC (admin, reviewer, and viewer roles with enforced permissions), tenant isolation, API key management, rate limiting, real-time webhook notifications, and the ability to approve or reject directly in Slack, Teams, and Telegram without opening the dashboard.
In terms of usage, SidClaw provides multiple integration paths. Users can try it locally: running npx sidclaw-demo will open the local governance dashboard at http://localhost:3030 with four preloaded scenarios built in (Claude Code's rm -rf, fintech transactions, DevOps scale-to-zero, and clinical lab orders), with no registration, Docker, or API key required. Integration options include: Claude Code Hooks (zero code, governing tool calls such as Bash, Write, Agent, and mcp__*); create-sidclaw-app (interactive scaffolding); MCP governance proxy (zero code, wrapping any MCP server); and SDK wrappers (one line of code per tool). It also provides a TypeScript SDK (npm) and a Python SDK (PyPI). For deployment, one-click deployment on Railway is recommended, which can start Postgres, the API, and the dashboard at the same time; Vercel can only host the Next.js dashboard and needs to be paired with a separately hosted Fastify API (Railway, Fly, Render, or Docker). Alternatively, users can start with the SDK using the free hosted API at app.sidclaw.com, without a platform.
Applicable scenarios include customer support (financial services), where policy flags and manual approval are required before an AI agent sends customer emails; DevOps infrastructure automation, where high-risk deployments require manual approval while read-only monitoring is allowed immediately; and healthcare clinical decision support, where lab orders are reviewed by physicians while prescriptions are prohibited by policy. The tool also targets security and compliance teams, mapping to regulatory requirements in the United States, the European Union, Switzerland, Singapore, and elsewhere.
In terms of limitations, the materials do not provide specific pricing or fee information; part of the platform uses the Functional Source License 1.1, which is source-available and may not be offered as a competing hosted service, and converts to Apache 2.0 after two years (March 2028); the SDK uses the Apache 2.0 license. The materials indicate that policy evaluation overhead is under 50 milliseconds. Other detailed information regarding performance, fees, and so on is not yet provided.
