Claude Desktop can use MCP servers to access files, repositories, databases, browser tools, and other local or remote capabilities. The safest setup starts with a narrow workflow and one server, then expands only after testing.
This guide focuses on practical setup decisions: choosing the first server, writing placeholder-based config, restarting Claude Desktop, testing discovery, and reviewing permissions before sensitive data is connected.
Key takeaways
- Start Claude Desktop MCP with one low-risk server and one sample workflow.
- Use placeholder examples for tokens and paths in any shared notes.
- Review permissions before enabling broad filesystem, GitHub, database, or browser access.
Plan the Claude workflow first
Decide what Claude should do before adding tools. A writing assistant may only need a docs folder. A coding assistant may need a project folder and GitHub. A data assistant may need read-only database access after sandbox testing.
This workflow-first approach prevents a common MCP mistake: enabling broad capabilities without a task or review model.
- Docs and knowledge-base workflows
- Repository onboarding
- Issue and PR research
- Read-only data exploration
- Browser research with isolated permissions
Create a Claude Desktop mcpServers entry
Claude Desktop setups commonly define server entries with command, args, and environment values. Keep public examples generic and put real values only where the local client can read them privately.
After saving config, fully restart Claude Desktop. If the server does not appear, check the command, package install behavior, file paths, and environment variable names.
- Use placeholder env names
- Avoid committing local Claude config
- Restart after every config change
- Test discovery before adding more servers
Good first Claude MCP servers
A good first server has low setup friction and low blast radius. Filesystem access scoped to a sample folder, GitHub access with limited token scopes, or a browser/search server for public docs can work well.
For databases and team communication, start with sandbox resources and read-only permissions because mistakes can expose private data or trigger external effects.
- Filesystem for selected folders
- GitHub for repository context
- Browser for public research
- Database only with read-only sandbox data
- Slack only with clear channel boundaries
Verify and secure the setup
Ask Claude to perform a harmless read, summarize what tools it can see, and explain what data it can access. Then document approvals for any write-capable actions.
Before production use, review prompts, logs, errors, credential storage, and rollback steps. Treat external files and web pages as untrusted content.
- Run a harmless read test
- Confirm no secrets appear in logs
- Document server purpose
- Require approval for writes
- Remove stale servers
Implementation checklist
- Choose a Claude workflow
- Add one server entry
- Use placeholder-only shared examples
- Restart Claude Desktop
- Run a harmless discovery test
- Review permissions before sensitive data
FAQ
How do I set up MCP in Claude Desktop?
Choose a server, add it to Claude Desktop's MCP config using the client-supported format, restart Claude Desktop, then run a harmless test call.
Which MCP server should Claude users install first?
A narrowly scoped filesystem server or a low-risk public research server is often a good first test.
Can Claude Desktop use GitHub MCP servers?
Yes, if the server is configured correctly and the token permissions match the intended workflow.
Should Claude Desktop config include real tokens in examples?
No. Shared examples should use placeholders. Real values belong only in private local settings or a secret manager.
Why is my Claude MCP server not showing up?
Common causes include invalid command paths, missing packages, malformed JSON, missing environment variables, or not restarting Claude Desktop after editing config.