DEV Community

Cover image for Introducing SidClaw: The Missing Approval Layer for AI Agents
SidClaw
SidClaw

Posted on

Introducing SidClaw: The Missing Approval Layer for AI Agents

#security #ai #opensource #agents

Your AI agents can send emails, execute trades, scale infrastructure, and order lab tests. But who approved that? Who saw what the agent was about to do before it did it? Who has the audit trail?

Today we're launching SidClaw — an open-source governance platform that adds the missing approval step to AI agent workflows.

The problem nobody else is solving

The AI agent governance space is heating up. Okta is working on identity for agents. SailPoint has machine identity. WorkOS ships authorization primitives. They all handle the same three things: identity, policy, and audit.

But none of them ship the approval step — the part where a human sees rich context about what an agent wants to do, reviews the risk level, sees the agent's reasoning, and makes an informed decision before the action executes.

That approval step is exactly what regulators are asking for:

  • FINRA's 2026 Regulatory Oversight Report explicitly calls for "human in the loop" agent oversight and "explicit human checkpoints before execution" for AI agents that act or transact.
  • The EU AI Act (enforceable August 2, 2026) requires human oversight, automatic logging, and risk management for high-risk AI systems. Penalties reach up to €35 million or 7% of global annual turnover.

We saw that the gap is obvious and that the tooling to fill it didn't exist.

How SidClaw works

Four primitives govern every agent action:

Agent wants to act → SidClaw evaluates → Policy decides → Human approves (if needed) → Action executes → Trace recorded
Enter fullscreen mode Exit fullscreen mode
  1. Identity — Every agent is registered with an owner and scoped permissions
  2. Policy — Every action is evaluated against explicit rules (allow / approval_required / deny)
  3. Approval — High-risk actions surface a context-rich card to a human reviewer: what the agent wants to do, why it was flagged, the agent's reasoning, and the risk level. One-click approve or deny.
  4. Trace — Every decision creates a tamper-proof audit trail with hash-chain integrity

The SDK wraps your existing tools. No changes to your agent logic:

import { AgentIdentityClient, withGovernance } from '@sidclaw/sdk';

const client = new AgentIdentityClient({
  apiKey: process.env.SIDCLAW_API_KEY,
  apiUrl: 'https://api.sidclaw.com',
  agentId: 'your-agent-id',
});

// Wrap any async function with governance
const sendEmail = withGovernance(client, {
  operation: 'send_email',
  data_classification: 'confidential',
}, async (to, subject, body) => {
  await emailService.send({ to, subject, body });
});

// Now this call is governed:
// - Policy says "allow"? Executes immediately.
// - Policy says "approval_required"? Waits for human approval.
// - Policy says "deny"? Throws ActionDeniedError. No email sent.
await sendEmail('customer@example.com', 'Follow-up', 'Hello...');
Enter fullscreen mode Exit fullscreen mode

Or in Python:

from sidclaw import SidClaw
from sidclaw.middleware.generic import with_governance, GovernanceConfig

client = SidClaw(api_key="ai_...", agent_id="your-agent-id")

@with_governance(client, GovernanceConfig(
    operation="send_email",
    data_classification="confidential",
))
def send_email(to, subject, body):
    email_service.send(to=to, subject=subject, body=body)
Enter fullscreen mode Exit fullscreen mode

13 framework integrations

SidClaw wraps whatever you're already using:

  • LangChain / LangGraph — governTools() wraps tool arrays
  • OpenAI Agents SDK — governOpenAITool() wraps function tools
  • Vercel AI SDK — governVercelTools() wraps tool objects
  • CrewAI — governCrewAITool() wraps task tools
  • Pydantic AI — governance_dependency() for tool functions
  • MCP — GovernanceMCPServer proxy wraps any MCP server
  • Composio, Claude Agent SDK, Google ADK, LlamaIndex — built-in support
  • Slack & Telegram — approve/deny directly from chat
  • GitHub Actions — governance checks in CI pipelines

Both TypeScript (npm) and Python (PyPI).

Try it right now

The fastest way — 60 seconds:

npx create-sidclaw-app my-agent
cd my-agent
npm start
Enter fullscreen mode Exit fullscreen mode

This creates a working governed agent with 3 demo tools:

  • search_docs — allowed instantly
  • send_email — requires your approval
  • export_data — blocked by policy

Or try our live demos without installing anything:

Each demo uses real SidClaw governance. The policy evaluation, approval workflow, and audit traces are 100% authentic.

The backstory: built in one session

Here's the part that might interest the AI/dev community: the entire SidClaw platform — API, dashboard, documentation site, landing page, three vertical demos, TypeScript SDK, Python SDK, 13 integrations, 705 tests, production deployment — was built with tenant isolation, hash-chain audit trails, RBAC, Stripe billing, and one-click deploy. We went from market research to live production fast!

What's next

SidClaw is free during early access. The Apache 2.0 SDK is open-source forever. The platform has visible pricing (Free / Starter CHF 199/mo / Business CHF 999/mo / Enterprise) but everything is free right now.

We're looking for design partners — especially teams in financial services, healthcare, and platform engineering who are deploying AI agents and need governance before the August 2026 EU AI Act deadline.

Links:

We'd love your feedback. Try a demo. Star the repo. Tell us what's missing.

Top comments (0)