Human-Controlled

You run a Human-Controlled MCP Server that lets Claude delegate real-time decisions and information gathering to you through a live web dashboard. Claude sends requests, and you respond via a responsive interface. This setup gives you hands-on control in situations that need human judgment, while keeping the rest of Claude’s workflow fast and seamless.

How to use

Connect Claude to your web dashboard via the MCP protocol, then respond to requests in real time. When Claude needs a human decision, a search, or a direct question answered, the MCP client forwards the request to your local server, and your web interface displays the task. You can then submit your reply, which Claude will incorporate back into its reasoning.

You will see three core tools available to Claude: ask_human to pose a question, human_search to request information gathering, and human_decision to deliberate on options. Each request appears in the dashboard with the relevant context, so you can provide a precise and timely response.

To complete a request, open the web interface at http://localhost:5001, read the incoming task, and enter your answer. The timeout for pending requests is five minutes, after which the system will handle the timeout automatically.

How to install

Prerequisites you need before starting: Python 3.x and pip for dependency management.

pip install -r requirements.txt

Configure Claude Desktop to connect to your local MCP server by editing your Claude Desktop configuration file. Use the following structure and replace the path with the absolute location of your Python script.

{
  "mcpServers": {
    "human-controlled": {
      "command": "python",
      "args": [
        "/absolute/path/to/human_mcp_server.py"
      ]
    }
  }
}

Start the server using either of these options.

# Option A: Via Claude Desktop
# Restart Claude Desktop after configuring
# Open the web interface at http://localhost:5001

# Option B: Standalone testing
python human_mcp_server.py

Configuration and access details

The web interface runs an MCP-enabled control panel on port 5001. It is a local web UI that displays incoming Claude requests and allows you to submit responses. You can access it using http://localhost:5001 or http://127.0.0.1:5001.

If you need to adjust the port, modify the startup settings in your server script so the web server binds to a different port, for example port 8080.

Security and maintenance notes

This server binds to all interfaces by default, which makes it accessible on your local network. If you only need localhost access, bind to 127.0.0.1. Consider adding authentication if you expose the interface beyond your machine.

If you encounter a connection issue, verify that Claude Desktop has a valid absolute path to your script, restart Claude Desktop after changes, and check your logs for errors.

If the web interface becomes inaccessible, ensure the server is running and that port 5001 is not blocked by a firewall. You can test port availability with system tools and confirm the interface by visiting http://127.0.0.1:5001.

Troubleshooting tips

Server won’t start: make sure port 5001 isn’t in use and all Python dependencies are installed with pip.

Claude can’t connect: ensure the path in the Claude Desktop configuration is an absolute path and restart Claude Desktop after changes.

Web interface not accessible: confirm the server is running and accessible at http://127.0.0.1:5001, then check firewall settings.

Requests timing out: the default timeout is five minutes; if needed, you can raise max_wait in the server code and verify the web interface remains open to poll for results.

Notes and tips

The setup leverages a Python-based MCP server paired with a real-time web dashboard. You can respond to Claude’s requests through a clean, minimal interface, enabling precise human inputs without changing Claude’s core workflows.

Available tools

ask_human

Pose a question to the human operator and wait for a response through the web interface.

human_search

Request information gathering from the human operator via the dashboard.

human_decision

Ask the human to choose between options or provide a recommendation.