cQnce MCP

@cqnce/mcp-server

Official MCP server for cQnce — add human-in-the-loop authorization to any AI agent or workflow.

Connect Claude, Cursor, GitHub Copilot, or any MCP-compatible client to cQnce so the AI can request human approval before performing risky or irreversible actions.

Quick start

npx @cqnce/mcp-server

Set the CQNCE_API_KEY environment variable to your project API key before running.

Approval modes

Mode 1 — Agent-requested approval

The agent calls wait_for_approval or submit_authorization_request itself, as instructed in the system prompt or via tool discovery. This is the easiest integration path and works well for supervised workflows.

Risk: a compromised or misconfigured agent may simply not call the tool.

Mode 2 — Enforced approval gate (recommended for production)

The host application or executor intercepts every sensitive tool call _before_ it runs and requires a prior cQnce approval, regardless of what the agent requested. The gate lives outside the model context — the agent cannot skip it.

This is a real security boundary. Build enterprise and compliance-sensitive integrations on this mode.

# Example: host-side interception (framework-agnostic)

HIGH_RISK_TOOLS = {"send_payment", "deploy_production", "delete_customer"}

def execute_tool_call(tool_call): if tool_call.name not in HIGH_RISK_TOOLS: return run_tool(tool_call) decision = cqnce.submit_and_wait( payload={"action": tool_call.name, "parameters": tool_call.arguments}, timeout_seconds=300, ) if decision["status"] != "APPROVED": return {"error": "not_authorized", "status": decision["status"]} return run_tool(tool_call)

Add to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows):

{

"mcpServers": { "cqnce": { "command": "npx", "args": ["-y", "@cqnce/mcp-server"], "env": { "CQNCE_API_KEY": "your-project-api-key" } } } }

Cursor

Add to .cursor/mcp.json in your project (or the global ~/.cursor/mcp.json):

{

"mcpServers": { "cqnce": { "command": "npx", "args": ["-y", "@cqnce/mcp-server"], "env": { "CQNCE_API_KEY": "your-project-api-key" } } } }

Any MCP client (stdio transport)

CQNCE_API_KEY=your-project-api-key npx @cqnce/mcp-server

Cloud agents (Streamable HTTP / remote MCP)

For cloud-hosted agents that cannot run local processes, use the cQnce MCP Worker deployed on Cloudflare Workers. It speaks the MCP Streamable HTTP transport and is stateless — every request authenticates with your API key.

Configure your cloud agent framework to connect to:

https://mcp.cqnce.app/mcp

Authorization: Bearer

Example (Claude API with MCP):

import anthropic

client = anthropic.Anthropic()

response = client.beta.messages.create( model="claude-opus-4-5", max_tokens=1024, mcp_servers=[ { "type": "url", "url": "https://mcp.cqnce.app/mcp", "name": "cqnce", "authorization_token": "your-project-api-key", } ], messages=[{"role": "user", "content": "..."}], betas=["mcp-client-2025-04-04"], )

Self-hosting

The Worker source is in this repository. Deploy your own instance to Cloudflare Workers:

npm install

npx wrangler deploy

Set CQNCE_BASE_URL in the Cloudflare dashboard if you use a private cQnce deployment (defaults to https://api.cqnce.app).

Configuration

Tools

Core (requires CQNCE_API_KEY)

Admin (requires CQNCE_ADMIN_TOKEN)

Project management, routing rule configuration, agent/team management, and webhook callbacks.

Recommended system prompt

When integrating Claude via the Anthropic API, add this system prompt to enforce the authorization policy automatically — without relying on the user to request it each time:

You have access to the cQnce tool for human-in-the-loop authorization.

ALWAYS call wait_for_approval BEFORE performing any action that is:

• Destructive or irreversible (deleting data, dropping tables, removing files)
• Affecting production systems (deployments, database migrations, config changes)
• Financial (payments, transfers, subscription changes)
• Involving credentials or access control (creating/revoking API keys, changing permissions)

In the approval payload, include:

• "action": what you are about to do
• "target": what resource is affected
• "reason": why this action is needed
• "risk": what happens if approved or rejected
• "attachments": (optional) list of supporting files as { name, contentType, data (base64) } — include logs, diffs, screenshots, or any evidence that helps the reviewer decide

When the response status is "REJECTED":

• Read the rejection reason carefully.
• If the reason identifies missing information or context, gather that information, enrich the payload (or attachments), and resubmit via wait_for_approval.
• Only give up if the rejection reason makes clear the action itself is not permitted.

Proceed ONLY if the returned status is "APPROVED". If "EXPIRED" or the rejection reason leaves no actionable path forward, stop and explain.

Example

Once configured, you can tell Claude:

"Before deleting the production database, ask for human approval via cQnce."

Claude will call wait_for_approval with the action details, pause until a human approves or rejects from the cQnce mobile app, and only proceed if the status is APPROVED. If rejected with a reason (e.g. "missing rollback plan"), Claude will gather that information and resubmit automatically.

Requirements

• Node.js 18+
• A cQnce account and project API key — sign up free