Claude for DevOps: Connect Your Stack with MCP
Claude is a strong DevOps assistant for reasoning through incidents, reading configurations and planning changes. On its own, though, it only knows what you tell it. MCP (Model Context Protocol) connectors let Claude retrieve facts from your actual systems during a conversation. This article covers what a DevOps MCP setup should connect, the pitfalls, and how to set one up.
What Claude needs to know for DevOps work
DevOps questions rarely live in one tool:
| Question | Where the answer lives |
|---|---|
| "What depends on this database?" | Cloud infrastructure |
| "Is checkout erroring right now?" | Observability, error tracking |
| "Why did the deploy fail?" | CI/CD |
| "Where is the timeout configured?" | Code repositories |
You could add a separate MCP server for each tool, but then Claude has to stitch the answers together itself, and the same service often has a different name in every tool.
Pitfalls of a DIY DevOps MCP setup
- Write access by accident. Many tool integrations can make changes. For investigations, read-only is safer.
- Name mismatches. checkout-svc in AWS, checkout in Datadog and checkout-pipeline in Jenkins are the same thing to you, but not to Claude.
- Conflicting sources. When two tools disagree, an assistant might blend the numbers into a confident but wrong answer.
- False zeros. "No errors found" from a tool that isn't connected, or wasn't scanned, looks the same as a real zero.
One connector for the whole stack
Vörr connects to Claude through one read-only MCP connector and links code, cloud infrastructure, observability and CI behind it:
- Repositories: GitHub, GitLab or Bitbucket. Vörr reads committed code and never pushes or changes anything.
- Cloud: AWS, Azure and Google Cloud with read-only credentials, plus Prometheus and Nagios where applicable.
- Connectors: observability and metrics, error tracking, distributed tracing, logs and CI/CD. For example, Grafana uses a Viewer service-account token, and Jenkins uses an API token.
- Service maps: link the different names a service has across tools, so answers match the right service. Matches show as explicit, inferred or unmapped.
- Multi-source answers: when several tools can answer, Vörr reports each result separately with its source. It doesn't average conflicting values, and one failing tool doesn't hide results from the others.
Set it up in Claude
- In Vörr, open Chat and choose Connect with Claude to copy the MCP URL.
- In Claude, go to Settings → Connectors → Add custom connector, give it a name, paste the URL into the remote MCP server URL field and select Continue.
- Approve access on Vörr's consent screen. The authorization is tied to your user and organisation and doesn't expose your stored credentials to Claude.
- Check it works by asking about a service you know is connected. The answer should show your real resources, with Vörr as the source.
You need a Claude plan that supports custom connectors.
Questions to try
- "Which dependency creates the largest blast radius for checkout?"
- "Which recent deployment overlaps with the error increase?"
- "Which repository and file define this behaviour?"
FAQ
Can Claude change my infrastructure through Vörr?
No. Vörr is read-only by design.
What if a tool isn't connected?
Vörr reports the limitation instead of returning a false "nothing found". Connect the source to fill the gap.
Does it work with ChatGPT too?
Yes. The same Vörr connector works with both Claude and ChatGPT.