Model Context Protocol | The How | How to connect MCP Servers to Claude Desktop | CampusX
Based on CampusX's video on YouTube. If you like this content, support the original creators by watching, liking and subscribing to their content.
Claude Desktop can connect to MCP servers using either one-click connectors or manual JSON configuration, depending on whether a curated connector exists.
Briefing
The practical takeaway: Claude Desktop can connect to multiple MCP servers—both local and remote—either through one-click “connectors” (for common SaaS tools) or via manual JSON configuration (for custom or less common servers). That dual approach matters because it lets teams move fast with ready-made integrations while still retaining the freedom to run their own MCP servers when they need specialized capabilities.
The walkthrough starts by setting expectations for a three-part “How” series. Today’s focus is the “how” of MCP communication between a client and servers, using Claude Desktop as the client and existing MCP servers as the targets. The plan is to install Claude Desktop, then connect it to four MCP servers: two local servers and two remote servers. The local pair includes a File System MCP server (for manipulating specific directories on the machine) and a Manim MCP server (for generating math visualizations). The remote pair includes a Google Drive MCP server (for reading documents from Google Drive) and a Twitter/X MCP server (for searching and posting tweets).
A key concept introduced is how Claude Desktop connects to MCP servers. The default method uses a JSON configuration file inside the AI host (Claude Desktop), where server details are added manually. The alternative method uses “connectors,” a built-in Claude feature that links Claude Desktop to MCP servers automatically—without editing JSON. Connectors handle the behind-the-scenes work such as authentication, API keys, and secure sign-in flows, which reduces setup friction for non-technical users and improves consistency compared with manual configuration. The tradeoff is that connectors are curated and managed by Anthropic’s team, mainly around widely used SaaS tools. Because MCP is an open standard that allows anyone to build clients and servers, forcing every MCP server to rely on Anthropic-managed connectors would close the ecosystem and create a bottleneck.
The File System MCP server demo shows how connectors work in practice: after installing the connector and granting access to specific directories (not the entire computer), Claude Desktop asks for permission before using the tool. A search for PDFs on the desktop returns results only after the user allows access, and Claude can also write files—demonstrated by generating a Python file and then verifying it exists on disk. The emphasis is on safety: access is scoped to approved folders, and users can add more directories only if they choose.
Next comes the Manim MCP server, which is treated as a local server without a ready-made connector. Setup requires installing dependencies (Python, MCP, Manim), cloning the repository, and then manually wiring the server into Claude Desktop’s JSON config. The payoff is a workflow where a user writes a high-level English prompt for a math concept, Claude generates the Manim code, executes it, and returns a visualization video—similar to the style of the “3Blue1Brown” channel. A practical caveat appears: perfect rendering may require LaTeX, and missing LaTeX can degrade mathematical symbol output.
For remote tools, the Google Drive MCP server connects via a direct connector and supports read-only operations—summarizing a document from Drive is shown, but creating or editing files is not. The Twitter/X MCP server setup is more involved: it requires adding configuration into Claude Desktop’s JSON, supplying API keys and access tokens, and then testing both tweet search and posting. Posting may fail if the account lacks proper authorization or permissions, though reading can still work.
Finally, a Weather MCP server is configured as a remote-style integration using a code path that runs on a different environment (via UVX), and the demo ends with an API/tool error that appears machine-specific. The session closes with guidance on discovering additional MCP servers: search for “awesome MCP servers” to find a curated, updated list on GitHub, plus other marketplaces and listing sites for further exploration.
Cornell Notes
Claude Desktop can connect to MCP servers in two ways: one-click “connectors” for curated SaaS integrations, and manual JSON configuration for custom or less common servers. The tutorial demonstrates four integrations: a local File System server (scoped directory access with permission prompts), a local Manim server (manual setup to generate math visualization videos from English prompts), a remote Google Drive server (read-only document search and summarization), and a Twitter/X server (JSON setup with API keys/tokens for tweet search and posting). The key reason both connection methods exist is scalability and ecosystem openness: connectors are managed for popular tools, while JSON keeps MCP truly open for anyone to run their own servers.
Why do connectors exist, and what do they change compared with JSON configuration?
How does the File System MCP server keep access safe?
What makes the Manim MCP server different from the File System server in setup?
What limitations apply to the Google Drive MCP server integration?
What can go wrong with the Twitter/X MCP server when posting tweets?
How does the Weather MCP server illustrate the difference between local and remote-style execution?
Review Questions
- When should a user choose connectors over JSON configuration, and what tradeoffs come with each approach?
- Describe the safety model for the File System MCP server and how permission prompts fit into it.
- What steps are required to integrate a local MCP server without a connector, and what additional dependency might be needed for high-quality Manim math rendering?
Key Points
- 1
Claude Desktop can connect to MCP servers using either one-click connectors or manual JSON configuration, depending on whether a curated connector exists.
- 2
Connectors reduce setup complexity by handling authentication, API keys, and secure connection logic behind the scenes.
- 3
The File System MCP server requires explicit directory scoping and permission prompts before Claude can read or write within allowed folders.
- 4
Manim MCP server integration may require manual JSON wiring because no connector is provided, and high-quality math rendering can depend on LaTeX being installed.
- 5
Google Drive MCP integration is read-only, enabling document search and summarization but not file creation or editing.
- 6
Twitter/X MCP posting can fail due to authorization or permission restrictions even when tweet search works.
- 7
A GitHub-curated “awesome MCP servers” list and other marketplaces help users discover additional MCP servers to integrate.