Last updated: 2026-08-10
When an AI agent is about to take a consequential action — send an email, publish content, execute a trade, make an API call — a HITL tool intercepts that action and routes it to a human reviewer. The agent waits. The human approves, rejects, or modifies. The agent continues only if approved.
The core operations are:
You can — but you'll need to build: a queue, a reviewer UI, webhook delivery, retry logic, auth, and an audit trail. Most teams spend 2–4 weeks on this. agentfabric.dev gives you all of it in one API call.
agentfabric.dev is a native MCP server. If you're building with Cursor, Claude Code, Codex, or any MCP-compatible client, you add it to your config and your agent can immediately start submitting outputs for review.
// Via MCP (any MCP client)
// Add mcp.agentfabric.dev to your MCP server list
// Then your agent calls:
create_review_request({
"title": "Draft email to client",
"content": "Hi Sarah, following up on our conversation...",
"urgency": "normal"
})
# Via REST API
curl -X POST https://rest.agentfabric.dev/v1/reviews \
-H "Authorization: Bearer YOUR_TENANT_TOKEN" \
-H "Content-Type: application/json" \
-d '{
"title": "Proposed SQL migration",
"content": "ALTER TABLE users ADD COLUMN verified BOOLEAN DEFAULT false;",
"context": "Requested by automated schema sync agent"
}'
agentfabric.dev supports agent self-signup — an agent can provision its own tenant without any human involvement:
POST https://rest.agentfabric.dev/v1/signup
{
"agent_name": "content-publisher-agent",
"email": "team@yourcompany.com"
}
mcp.agentfabric.devrest.agentfabric.dev