What is MCP: How Claude Code connects to external tools and data
MCP (Model Context Protocol) is a protocol for connecting AI applications to external tools and data sources. For Claude Code, after you configure an MCP server, Claude can call the capabilities it provides as needed during a conversation, such as querying external data or performing specific actions. MCP does not automatically give Claude arbitrary system permissions: what it can do depends on the tools exposed by the server, the configured credentials, and the scope of access you grant.
You can think of an MCP connection as having three parts: Claude Code is the client, the MCP server provides tools, and tool calls pass requests and results through the session. Servers can connect through local process communication or through methods such as HTTP. The method used depends on how the server is deployed and its connection instructions.
For beginners, first understand one key point: installing MCP usually does not mean installing a Claude Code plugin. Instead, it means telling Claude Code how to start or connect to a server. Once configured, Claude Code can discover the tools the server provides and call them when needed.
Installing MCP for the first time: Configuration methods, scope, and connection verification
Before getting started, make sure you have installed Claude Code and can run claude in the terminal. When configuring it for the first time, prioritize the connection commands or configuration examples provided in the service documentation. Do not arbitrarily replace startup arguments, addresses, or authentication methods. Different servers may require different runtimes, environment variables, or authorization flows.
A common way to install is to add an MCP server via the command line. The example below shows the parameter structure; <启动命令>, <参数>, and <服务地址> are placeholders that should be replaced with the actual values from the corresponding server documentation.
# 添加通过本地进程启动的服务器
claude mcp add --transport stdio --scope user my-server -- <启动命令> <参数>
# 添加通过 HTTP 连接的服务器
claude mcp add --transport http --scope user my-server <服务地址>
--transport specifies the connection type; --scope specifies the configuration scope. Common scopes include:
| Scope | Use case |
|---|---|
local | Use only in the current project, without sharing the configuration with other project members |
project | Write the configuration to the project for collaborators to share; check for sensitive information before committing |
user | Use across multiple projects for the current user |
For first-time use, consider user to avoid putting personal credentials or local machine paths in project configuration. If you want to share connection definitions with your team, consider project, but provide keys securely through environment variables or similar means instead of committing them directly to the repository.
After adding a server, you can use the following command to check the configuration, then start a Claude Code session:
claude mcp list
claude mcp get my-server
In Claude Code, use /mcp to check the connection status. Some services require OAuth authorization; follow the prompts in the interface to complete authentication. Once the connection appears healthy, make a clearly scoped request that will not cause destructive actions to confirm that the corresponding tools can be discovered and called. If the server requires specific environment variables, configure them with command-line arguments when adding it. Refer to claude mcp add --help for the syntax supported by your current version.
Managing MCP configuration: View, edit, troubleshoot, and security considerations
When managing MCP, first confirm which configuration Claude Code is actually reading. Different scopes affect where configurations are stored and how they are shared; servers with the same name may also exist in different scopes. Use claude mcp list to view registered servers and claude mcp get <名称> to view the configuration for an individual entry. Then modify or re-add it in the intended scope to avoid changing a connection used by another project.
If a connection fails, troubleshoot in this order:
- Confirm that the server address, startup command, and arguments match the service documentation.
- Check whether the required runtime is installed and whether the current terminal can find the corresponding command.
- Check environment variables, access tokens, and authorization status. Do not paste secrets into public logs or commit records.
- Recheck the status and error messages in
/mcp, and confirm that Claude Code has reloaded the configuration. - If the server connects but its tools are unavailable, check whether the server actually exposes tools and whether the current request is appropriate for those tools.
MCP servers receive the requests sent to them, and some tools may also read files, modify data, or trigger external actions. Before adding a server, confirm that its source is trustworthy, its permissions match your expectations, and you understand which resources it can access. For file-access services, expose only the directories needed for the task whenever possible. For tools with write or execution capabilities, check the target and impact of an operation before calling them.
Uninstalling and cleanup: Remove server configuration and related credentials
When you no longer need a server, you can use:
claude mcp remove my-server
If entries with the same name exist in multiple scopes, first use claude mcp list to confirm which configuration you want to remove, then specify the scope using parameters supported by your current version of Claude Code. Do not assume the scope based on the name alone. After removal, check the list again and use /mcp in Claude Code to confirm that the connection is gone.
Removing the configuration does not necessarily delete the server software or clean up external credentials. If the server was installed through a package manager, container, or standalone program, uninstall it according to its own documentation. If you configured tokens, environment variables, or OAuth authorization, revoke or rotate the credentials according to the service provider's process. Also check the version control status of project-level configuration to ensure that files or sensitive information you no longer need have not been left in the repository.
How MCP uses context: Tool definitions, call results, and on-demand loading
MCP affects context mainly through two components: tool definitions and tool call results. Tool definitions usually include names, purposes, parameter descriptions, and other information. The more servers and tools there are, and the longer their descriptions, the more information Claude may need to process when deciding which capabilities are available. Data returned after calling a tool also enters the conversation context. If a call returns a large amount of logs, file contents, or records, it can significantly increase context usage even when the tool definition is short.
Whether tools are loaded on demand depends on the Claude Code version and its tool discovery mechanism. Some versions use deferred discovery or search when there are many tools, reducing the amount of tool information presented at the beginning. This does not mean that call results will not use context, nor does it mean that all MCP servers support the same mechanism. Refer to the actual behavior of your local version, and avoid assuming that “configured but unused” means absolutely no overhead.
You can use /context to observe context usage in the current session and check whether tool responses are too long. Compacting context or starting a new session may sometimes ease accumulated pressure, but neither replaces controlling tool output. More effective approaches include reducing unnecessary servers and tools, limiting query scope, and having tools return only the information needed to complete the task.
Practical usage tips: Choose tools, control output, and improve workflow efficiency
When using Claude Code with MCP for the first time, start with a server that has a clear purpose and a limited permission scope. Verify the connection first, then test a single read-only operation, and only then try capabilities that may write or modify data. This makes it easier to tell whether a problem comes from the configuration, authorization, or the tool itself.
Be clear about the scope of your request, too. For example, instead of just saying “check the data,” specify what to query, the time range, the fields you want, and whether write operations are allowed. For tasks that return a lot of content, you can ask for a summary first, request only specific fields, or query using pagination and filters. Whether this works depends on whether the server's tools provide the relevant parameters.
For routine maintenance, follow these practices:
- Configure only a few necessary servers for each work scenario; don't add lots of tools just because you “might need them later.”
- Keep personal configuration separate from shared team configuration to avoid including local paths and credentials in project files.
- For operations such as writing, deleting, or sending data externally, ask Claude to explain the plan first, then confirm before proceeding.
- If a connection fails, check the configuration and authorization first; don't repeatedly paste keys or reinstall blindly.
- Remove connections you no longer use after a task, and clean up expired tokens.
The value of MCP lies in connecting Claude Code to the right external capabilities, not in connecting as many as possible. Starting with trusted servers and minimal permissions, checking the results and context overhead, and then expanding gradually is usually the safer way to get started.