How to connect Windsurf's Cascade coding agent to outside tools with MCP

What the guide covers

As dev.to reported, a tutorial walks through connecting Codeium's Windsurf editor to external tools using the Model Context Protocol, or MCP. MCP is a standard way for an AI assistant to talk to outside software, such as a database or a search service, instead of relying only on what you paste into a chat.

Windsurf's built-in assistant is called Cascade. The guide describes it as an agent, meaning it can do more than suggest code. It reasons across several files, runs terminal commands and indexes a whole codebase. MCP widens its reach beyond the project folder it is working in.

The post is a how-to rather than news, and it comes from the team behind AIStoreHub, whose deeper write-up is linked at the end. That context matters when you reach the part about their command-line tool.

What changes when Cascade has MCP

Without MCP, a coding assistant knows only what is in your chat and what its own index of your files turns up. The guide gives three examples of what changes once servers are connected.

• Database checks: Cascade writes a SQL migration, which is a script that changes a database's structure. It then tests that script against a local PostgreSQL copy running in Docker and confirms column types, without you opening another app.
• Current documentation: through search or fetch servers, it can pull up-to-date reference pages. The guide says this beats the assistant inventing calls to package features that no longer exist.
• Issue tracking: it can look up your GitHub issues to confirm what a pull request was supposed to do before it refactors code.

The common thread is verification. The assistant checks a real system instead of guessing from memory.

Where the settings live

Windsurf manages all of this through one file called mcp_config.json. On macOS and Linux it sits at ~/.codeium/windsurf/mcp_config.json. Windows uses a matching .codeium\windsurf folder under the user profile. You can also reach it from inside the editor through Settings, then Cascade, then Model Context Protocol (MCP), then Open Config.

The file follows the standard MCP layout. At the top is a dictionary called mcpServers, and each server gets a name and a recipe for starting it.

The guide's example defines two servers. One is a local SQLite database server, launched with the uvx command and pointed at a database file on disk. The other is a GitHub server, launched with npx, which receives a personal access token through an env section.

That token works like a password. The guide shows a placeholder, but a real one should be kept out of anything you share or commit to a repository.

Using it, and the approval step

Once the file is saved, you open the Cascade panel with Cmd + L on Mac or Ctrl + L elsewhere. You then ask for something that needs outside context. The guide's example is asking Cascade to inspect a SQLite database's structure and write a migration that adds a status column to a users table.

Cascade lists the MCP tools it can use and asks for your approval before running them. That approval step is the main safeguard in the setup. Each tool call is a chance for the assistant to read or change something real, so it is worth reading the request before you accept.

The guide does not say how detailed these approval prompts are or whether you can pre-approve certain tools, so check Windsurf's own settings for that.

The one-command shortcut, and a caution

The guide's last section promotes a free, open-source tool called the AIStoreHub CLI. You install it with pip, then run commands such as aistorehub add sqlite --client windsurf or the same with github. According to the guide, the tool finds your config file, backs up what you have, and merges the new server in without overwriting the rest.

It saves you from hunting for a hidden folder and hand-editing JSON, a file format where one missing comma can break everything. The post says the servers it adds are verified, but it does not explain how that verification works. Since the guide comes from the tool's makers, read the project's code and documentation before letting it touch your config.

The same caution applies to any server you add by hand. Each one is a program that runs on your machine and may receive your credentials, so install only ones you trust.