Once your MCP server runs locally, the next milestone is getting Claude Desktop to actually talk to it so you can start exercising your tools with real prompts.
- Claude Desktop prefers absolute paths inside the claude_desktop_config.json file.
- A running tag next to your server name in the settings panel confirms it connected successfully.
- Every input and output gets logged, so sensitive data should be handled deliberately in any real deployment.
This lesson is a preview from our Building Your First MCP Server and Client Course Online. Enroll in a course for detailed lessons, live instructor support, and project-based training.
Wiring up Claude Desktop is the first test that your server can play nicely with a production-grade client. It also shows off one of the nicer properties of MCP: the same server can be used from any compatible client, so the work you do here applies just as well to ChatGPT, Cursor, and other tools later on.
Editing the Claude Desktop Config
Inside Claude Desktop, open settings, go to the developer section, and click edit config. A window opens with the claude_desktop_config.json file preselected. Double click to open it in an editor of your choice. This is the file Claude looks at on startup to decide which MCP servers to launch.
The structure is simple. Each server gets a name, a command (usually UV for Python-based servers), and a set of arguments that tell Claude exactly where the project folder lives and which Python file to execute. Absolute paths are required, because relative paths behave inconsistently at launch. A good shortcut is to right-click the file in your editor and choose copy path, which yields the absolute path in one step.
Adding a New Server Entry
If this is your first MCP server, paste the full block into the file. If you already have servers defined, add a comma after the previous entry and drop your new block in alongside it. Give it a clear name that you will recognize later, like spacex-dashboard or aurora-watcher.
Save the file, close it, then fully quit and reopen Claude Desktop. The app reads the config on launch, so a restart is what actually picks up your changes.
Confirming the Connection
After Claude restarts, head back into the developer settings. A successful server shows up with a running tag next to its name, along with the command and arguments it was started with. You can click through to view the server's log output, which shows FastMCP spinning up, the inputs it received, and the outputs it produced.
Those logs are useful for two reasons. First, they make debugging far easier: any error message can be copied straight into a chat and used to diagnose the problem. Second, they are a reminder that MCP traffic is visible. Anything that flows through a third-party server will be logged by that server, so personal data, user IDs, and payment information deserve extra care.
Running Your First Real Prompt
Open a new chat and ask a question that your server should answer. For a SpaceX demo, asking about the last few launches is a good smoke test. Because Claude will naturally try to use the web for current events, it helps to steer it explicitly: use the dashboard MCP tool, and do not search the web. That forces the query through your server so you can watch it in action.
Claude may also ask permission to use a new tool the first time. Approve it, and the run continues. From there you can see which tools were invoked, how they were chained, and how the final response was assembled. Complex questions like listing the most frequent crew members across missions become possible because your outcome-based tools are doing the heavy lifting before the data ever reaches Claude.
Connecting a local MCP server to Claude Desktop comes down to four steps: edit the config, point it at absolute paths, restart the app, and confirm the running tag. Once that is in place, every new prompt becomes a live test of your tool design, and you can start iterating on the server with real feedback instead of guesswork.