DEV Community

Tomáš Grasl
Tomáš Grasl

Posted on

I Built an MCP Server for OpenClaw — Now My AIs Talk to Each Other

#mcp #automation #opensource #ai

There's been a lot of hype around OpenClaw (a.k.a. Claw bot) lately. At first, it seemed like a fun toy to play with — nothing more. But over time, I realized it could actually be useful, so I installed it on my own physical server and started experimenting.

The Problem with Chat Channels

OpenClaw supports communication through Discord and Telegram, but honestly, that felt limiting. Chatting with an AI agent through a messaging app takes away a lot of its potential. So I asked myself: what if I built an MCP server and connected it directly to Claude?

So I Built It

I created an MCP (Model Context Protocol) server for OpenClaw — check it out on GitHub — and I have to say, watching one AI communicate with another AI is both hilarious and surprisingly effective. What started as a fun experiment turned out to be incredibly powerful.

Real-World Use Cases

Here's where it gets interesting. Let me walk you through two workflows I'm actually using.

Auto-Fixing Bugs from Sentry

Imagine you have an n8n automation that reacts to bugs coming from Sentry. Here's the flow:

  1. A bug comes in from Sentry
  2. The n8n workflow picks it up and an AI node evaluates it — yes, this should be fixed
  3. Via MCP, it creates a prompt for OpenClaw
  4. Claw clones the project repository on the server
  5. It spins up Claude Code, which fixes the bug
  6. Claw creates a pull request on GitHub
  7. It reports back to the automation with the PR link
  8. You get a notification via Pushover or Slack

And yes, this actually works.

Jira Task Management with Claude Code

Here's another one. You have the OpenClaw MCP connected to Claude Code on your localhost. You have a task in Jira, but you don't want to load Claude Code with a dozen different MCP servers — that eats up your context window. Instead, you only have the OpenClaw MCP.

You ask Claw to look up the Jira task and send back what needs to be done. It returns the task as a ready-to-go prompt for Claude Code, and you can start working immediately. Once you're done, you notify Jira through Claw that the task is complete. Clean and simple.

A Word of Caution

OpenClaw is powerful, but it can also be dangerous if you don't configure it properly. It's gone off the rails on me a few times already. Make sure you set clear boundaries and review what it's doing, especially when it has access to your repositories and external services.

As an experiment and a productivity booster? Absolutely worth it. Just keep it on a leash. 🦞

Top comments (0)