Claude MCP Approval Server

The Claude MCP Approval Server is a lightweight, FastMCP‑based service that adds an extra layer of security to AI workflows by routing potentially sensitive tool calls through a human‑in‑the‑loop approval process. Instead of letting an assistant execute commands automatically, the server intercepts each tool invocation and sends a concise, actionable prompt to a designated WhatsApp number via Twilio. The user can then grant or deny the request with a single tap, ensuring that no unauthorized actions are taken without explicit consent.

This approach addresses the growing concern around autonomous AI execution in production environments. Developers who rely on Claude to interact with external APIs, manipulate files, or trigger infrastructure changes often need a safeguard that can be accessed from any device. By leveraging WhatsApp’s ubiquity and Twilio’s robust messaging infrastructure, the server delivers approvals in real time without requiring a custom UI or additional authentication layers. The quick‑reply buttons provide an intuitive, mobile‑friendly interface that can be used from anywhere in the world.

Key capabilities of the server include:

• Permission Prompts: Every tool call is intercepted and reformatted into a readable request that includes the command, its arguments, and the reason for execution.
• WhatsApp Integration: Twilio’s API sends the prompt as a message with “Approve” and “Deny” buttons, enabling instant human response.
• Auto‑Expiry: Requests automatically expire after five minutes to prevent stale approvals and reduce security risk.
• Database Tracking: An SQLite database records each request, its status, and timestamps for auditability.
• FastMCP Compatibility: The server exposes a standard MCP endpoint () that Claude can call directly, making setup seamless for existing MCP‑enabled assistants.

Real‑world scenarios where this server shines include automated deployment pipelines, data‑scraping operations that require compliance checks, or any environment where an AI assistant interacts with financial systems. By requiring a human approval step before execution, teams can maintain governance while still enjoying the speed and flexibility of AI‑driven automation.

Integrating the approval server into an AI workflow is straightforward: configure Claude to use the flag pointing at the server’s MCP endpoint, expose the service via ngrok or Tailscale, and set Twilio webhooks to forward incoming messages. Once operational, every tool call triggers a WhatsApp message; the user’s tap immediately resolves the request, allowing Claude to proceed or abort. This tight loop preserves developer productivity while upholding stringent security and compliance standards.