SidClaw

SidClaw is a governance platform for AI agents. It answers a question that every team deploying autonomous agents eventually faces: how do you stay in control?

SidClaw sits between your agent and the tools it calls. Every action is evaluated against policies you define, and when something is sensitive or high-risk, a human reviewer sees exactly what the agent wants to do — with full context — and approves or denies it before it happens.

Four Primitives

SidClaw is built on four primitives that form a chain: Identity, Policy, Approval, Trace.

Identity — Every agent gets a registered identity with an owner, team, environment, authority model, and autonomy tier. Identity determines what policies apply and what level of oversight is required. Learn more

Policy — Rules that evaluate agent actions. Each policy matches on operation, target integration, resource scope, and data classification, then returns one of three effects: allow, approval_required, or deny. Higher-priority policies take precedence. Learn more

Approval — The core differentiator. When a policy evaluates to approval_required, SidClaw creates a context-rich approval card showing what the agent wants to do, why it was flagged, the risk classification, and the agent's reasoning. A human reviewer approves or denies. Separation of duties ensures agent owners cannot approve their own agent's requests. Learn more

Trace — Every evaluated action produces a tamper-proof audit trail. Each trace is a chronological chain of events — from initiation through identity resolution, policy evaluation, approval (if needed), and final outcome. SHA-256 hash chains guarantee integrity. Learn more

How It Works

Agent calls tool → SDK evaluates action → Policy Engine decides
                                            ├─ allow → execute immediately
                                            ├─ approval_required → human reviews → approve/deny
                                            └─ deny → blocked, ActionDeniedError thrown

The SDK's withGovernance() wrapper handles the entire flow. Your agent code calls the wrapped function normally. If approval is required, the SDK automatically polls until a reviewer decides.

Integrations

SidClaw provides first-class integration with the major agent frameworks:

• MCP (Model Context Protocol) — Governance proxy server that wraps any MCP tool server
• LangChain — governTool() wraps LangChain StructuredTool instances
• Vercel AI SDK — governVercelTool() wraps Vercel AI tool() definitions
• OpenAI Agents SDK — governOpenAITool() wraps OpenAI function tools
• CrewAI — governCrewAITool() wraps CrewAI tool definitions
• Generic — withGovernance() wraps any async function

Get Started

• Quick Start — Install the SDK, wrap your first tool, and see governance in action in under 2 minutes.
• Concepts — Understand the four primitives in depth.