Fivetran MCP Server

The Fivetran MCP Server connects AI assistants such as Claude Code or Cursor to your Fivetran account. Fivetran is a service for automated data integration: it copies data from sources such as databases and business applications into a data warehouse and keeps it current. In day-to-day work the MCP server answers typical operations questions, for example whether the latest sync of a Postgres source succeeded or whether any connection in the account is broken. Instead of clicking through the dashboard you ask the question in chat and get an answer based on the Fivetran API. The project lives in the GitHub organization fivetran, is released under the MIT license, currently has around 25 stars, is not archived, and was last updated on October 21, 2026. The README does not explicitly call the project official; however, it is maintained by Fivetran itself in its GitHub organization.

What can the Fivetran MCP Server do?

According to the README the server works with three kinds of tools. list_endpoints discovers API endpoints by category or search term, get_schema returns the parameters and schema of an endpoint, and tools following the pattern <resource>_<action>, for example connections_read or connections_write, execute the endpoints. Which of them are visible depends on your scope and an optional denylist. This design keeps the assistant's context lean: it does not load the whole API up front but explores it as needed.

Read-only by default

An important difference from many other servers lies in the default. The environment variable FIVETRAN_SCOPE sets the ceiling for permissions: read is the default, alongside read/write and read/write/delete. With DISALLOWED_ACTIONS you additionally block individual actions or endpoints. The README points out that write and delete actions only carry advisory confirmation warnings; the server does not enforce confirmation. For endpoints concerning credentials, such as user API keys and system keys, it explicitly recommends a denylist.

Local or as an HTTP server

By default the server runs locally over stdio on your machine. The recommended installation uses uvx --from git+https://github.com/fivetran/fivetran-mcp fivetran-mcp. The README warns that a bare uvx fivetran-mcp installs a different, unrelated package from the Python package index. Alternatively the server can run as a streamable HTTP server that several clients reach over the network. In this mode the credentials come from the Authorization header of each request; according to the README, OAuth support is so far only scaffolding and not yet functional.

Who is the Fivetran MCP Server for?

It is worthwhile for data engineering teams and analysts who use Fivetran in production and want to settle operations questions faster: which connections are paused, where did a sync fail, how is a connector configured? Without a Fivetran account it adds nothing. Start with the default scope read, enable write permissions only once you know the flow, and block sensitive endpoints.

Costs

The MCP server itself is free under the MIT license. A Fivetran account with an API key and secret is required. According to the pricing page there is a free plan as an entry point for low data volumes as well as paid plans with usage-based billing; for large projects Fivetran offers enterprise agreements through sales. Current terms are on Fivetran's official pricing page.

Fivetran account with an API key and API secret, uv (to start via uvx), and an MCP client with stdio support. For HTTP mode, a reachable host.

With the Fivetran MCP Server an AI assistant queries the state of your data pipelines. This guide leads from credentials to the first question.

Step 1: Prerequisites

You need a Fivetran account with an API key and API secret, the tool uv, and an MCP client such as Claude Code or Cursor. You create the key in your Fivetran account; ideally issue it for a dedicated account with limited permissions.

Step 2: Set up the server

In Claude Code the command is claude mcp add fivetran --env FIVETRAN_API_KEY=<key> --env FIVETRAN_API_SECRET=<secret> -- uvx --from git+https://github.com/fivetran/fivetran-mcp fivetran-mcp. Use exactly this form: according to the README a bare uvx fivetran-mcp installs a different, unrelated package.

claude mcp add fivetran --env FIVETRAN_API_KEY=<key> --env FIVETRAN_API_SECRET=<secret> -- uvx --from git+https://github.com/fivetran/fivetran-mcp fivetran-mcp
uvx fivetran-mcp

Scope and denylist

The default scope FIVETRAN_SCOPE=read allows read access only. You enable write and delete permissions with read/write or read/write/delete. The confirmation warnings on write actions are advisory only. So use DISALLOWED_ACTIONS to block endpoints concerning credentials, such as user API keys and system keys.

Step 3: Run as an HTTP server

For teams the server can be started as a streamable HTTP server. Then the credentials come from the Authorization header of each request. According to the README, OAuth support is not yet functional.

First questions

Start with reads: "Did the latest syncs of my Postgres source succeed?" or "Are there broken connections in my account?" Only then have configuration changed, and only with an elevated scope.

Fivetran API key and secret through the environment variables FIVETRAN_API_KEY and FIVETRAN_API_SECRET; in HTTP mode through the Authorization header of the request.

Access to the account's Fivetran API within the configured scope: read-only by default, optionally with write and delete permissions; individual actions can be excluded through a denylist.

Retrieved configuration and status data enter the AI client context and therefore reach the model provider. Processing by Fivetran follows the provider's privacy policy.

Unintended changes or deletion of connections with an elevated scope (confirmations are advisory only), exposure of credentials via unblocked key endpoints, an overly broad API key, and prompt injection via content from the account.