We Built an MCP Server So AI Agents Can Manage Your Data Pipelines

The problem with switching between dashboards and code

If you work with data pipelines, you know the drill. You are deep in your codebase, building a new integration for a client. You need to check the schema definition. So you leave your editor, open the FileFeed dashboard, navigate to Schemas, find the right one, read the field definitions, switch back to your editor. Then you realize you need to look up the pipeline mappings too. Back to the dashboard.

This context switching is not just annoying. It breaks flow state. Studies on developer productivity consistently show that every context switch costs 10 to 15 minutes of recovery time. If you are switching between your IDE and a dashboard five times during a coding session, you have lost an hour to navigation alone.

We wanted to fix this. Not with another browser tab or a CLI tool you have to memorize commands for. We wanted FileFeed operations to live where engineers already work: inside their AI coding assistant.

What is MCP and why it matters

The Model Context Protocol is an open standard created by Anthropic that lets AI assistants connect to external tools and data sources. Think of it as a universal adapter between large language models and the services they need to interact with. Instead of each AI tool building custom integrations for every platform, MCP provides a shared protocol that works across all of them.

In practice, this means you can install a single MCP server and immediately use it from Claude Desktop, Cursor, VS Code with Copilot, Windsurf, or any other tool that supports the protocol. The AI assistant discovers the available tools automatically, understands what they do, and can call them on your behalf.

We shipped the @filefeed/mcp-server package because it seemed obvious: if AI assistants can write code, they should also be able to manage the data infrastructure that code depends on. One command, and your assistant gains full access to FileFeed.

What the FileFeed MCP server can do

The server exposes 38 tools across six resource groups. Every operation you can perform in the FileFeed dashboard or through the REST API is available as an MCP tool. This is not a limited subset. It is the full API surface.

Clients

Create, update, list, retrieve, and delete SFTP clients. Test connections to verify SFTP credentials are working. When you ask your assistant to set up a new client for a partner, it creates the client, provisions SFTP credentials, and hands you the connection details all without leaving your editor.

Schemas

Define the target data structure for your pipelines. Create schemas with field names, types, and validation rules. Validate sample data against a schema before you go live. If you are building an integration and need to verify that a client's data matches expectations, your assistant can run the validation tool and tell you exactly which fields fail.

Pipelines

Create pipelines that connect clients to schemas with field mappings and transformations. Toggle pipelines on and off. Update mappings when a client changes their export format. The assistant understands the relationship between clients, schemas, and pipelines, so you can say "create a pipeline for ACME that maps their export to the employee schema" and it handles the rest.

Pipeline Runs

List runs filtered by status, client, or pipeline. Retrieve processed data with pagination. Acknowledge completed runs. Reprocess failed runs. Get presigned URLs for downloading original or processed files. This is where the day-to-day operations happen, and having it accessible from your assistant means you can debug a failed run without ever opening the dashboard.

Webhooks and Outbound

Manage webhook endpoints for event notifications. Create outbound file export sessions for pushing data into pipelines programmatically. The outbound tools support chunked multipart uploads, so your assistant can push large datasets into FileFeed in parts.

Setting it up takes two minutes

You need two things: an API key and one configuration entry. Get your API key from the FileFeed dashboard under My Account → Security Settings. Then add the server to your AI tool's configuration.

Claude Desktop

Open your Claude Desktop config file and add the following to the mcpServers section:

{
  "mcpServers": {
    "filefeed": {
      "command": "npx",
      "args": ["@filefeed/mcp-server"],
      "env": {
        "FILEFEED_API_KEY": "your-api-key-here"
      }
    }
  }
}

Restart Claude Desktop. You will see the FileFeed tools appear in the tool list. That is it.

Cursor and VS Code

The same configuration pattern works in Cursor and VS Code with GitHub Copilot. Add the MCP server to your workspace or global settings, point it at your API key, and the tools are available in your next chat session. The exact location of the config file varies by tool, but the server entry is identical.

Real workflows that become one-liners

The value of MCP tools is not that they replace the dashboard. It is that they remove the gap between thinking and doing. Here are real examples from our own team's daily workflow.

• "Show me all failed pipeline runs from the last 3 days." The assistant calls the pipeline runs list tool with a status filter, formats the results as a table, and tells you which clients and files failed. You can immediately ask it to reprocess the failed ones.
• "Create a new client called Acme Corp and set up a pipeline mapping their export to our employee schema." The assistant creates the client, looks up the schema, builds the pipeline with field mappings, and reports the SFTP credentials for the client to connect.
• "Check if run abc-123 processed correctly and show me the first 10 rows of data." The assistant retrieves the run details, checks the status, and fetches the first page of processed data. If something looks wrong, you ask it to reprocess.
• "How many pipeline runs completed for Samsung this week?" The assistant filters runs by client name, date range, and completed status. One question, one answer, no dashboard.

How we use it internally

We dogfood the MCP server every day. When we are developing new features in the FileFeed backend, we use the MCP tools to create test clients, schemas, and pipelines without writing setup scripts. When we are debugging a customer issue, we pull pipeline run data and check error messages directly from Cursor. When we want to validate that a schema change works correctly, we ask the assistant to run sample data against the schema and show us the validation results.

The biggest surprise was how much it changed support workflows. Instead of asking a customer to send screenshots of their pipeline configuration, our team can pull the configuration directly, compare it against what the documentation says, and identify mismatches in seconds. The MCP server turned FileFeed operations from a browser-dependent workflow into something that lives in the tool we are already using eight hours a day.

Open source and extensible

The @filefeed/mcp-server package is published on npm and the source is open. If you want to see how we implemented a specific tool, you can read the code. If you want to extend it with custom tools specific to your workflow, you can fork it.

We chose MCP over a proprietary integration because we believe the tooling ecosystem is moving toward open protocols. Today it is Claude and Cursor. Tomorrow it will be every IDE and every AI assistant. Building on MCP means the work we did once benefits every tool that adopts the standard.

If you are building B2B data pipelines and your team uses AI coding assistants, the MCP server is worth the two minutes it takes to install. For a look at how the FileFeed AI Assistant brings the same capabilities into the dashboard, see our product walkthrough. You will wonder how you worked without it.

Read the full MCP server documentation or install it now with npx @filefeed/mcp-server. If you do not have a FileFeed account yet, sign up for free and get your API key to start.