Fivetran MCP

What does a Fivetran MCP server do?

A Fivetran MCP server lets an AI assistant check your pipelines or query the data they load. fivetran-mcp answers questions such as which connections are failing, and can start syncs if you allow writes. Agent Context MCP gives assistants your Context Layer and read-only SQL on your destination warehouse.

• Which to use: sync health and connection changes need fivetran-mcp. Business questions on warehouse data need Agent Context MCP, which lists no connection or sync tools.
• Context Layer is a separate Fivetran product, also in Public Preview. It writes an AGENTS schema to your destination from the datasets you select, and can add documents, systems of record and analytics sources. A build typically takes about a day.

Is there an official Fivetran MCP server?

Yes. Fivetran publishes both. Agent Context MCP is in Fivetran's docs, labelled Public Preview. fivetran-mcp sits in Fivetran's GitHub org under an MIT license, and its README states no support terms. A November 2025 Fivetran post says Fivetran has no officially supported MCP server, so ask Fivetran about support before you depend on it.

• Fivetran's fivetran/skills plugin bundles fivetran-mcp for Claude Code.
• Other GitHub repos use the Fivetran name from outside Fivetran's org, so check the owner before you install one.

What is the Fivetran MCP server URL, and how do you sign in?

Agent Context MCP lives at https://api.fivetran.ai/mcp. Fivetran's guide sets ChatGPT to OAuth, and custom agents send a system key as a bearer token. fivetran-mcp has no public URL. It runs on your side with a Fivetran API key and secret, so you create a key in the dashboard first.

• Run fivetran-mcp with uvx --from git+https://github.com/fivetran/fivetran-mcp fivetran-mcp. Bare uvx fivetran-mcp doesn't work, because those PyPI names belong to unrelated projects.
• A scoped API key inherits your RBAC permissions, so a restricted Fivetran user limits the server too. The REST API also has system keys, which aren't tied to a user and set READ, MANAGE or NONE per permission.
• fivetran-mcp can also run as an HTTP service, with the key and secret in a Basic Authorization header on each request. Its OAuth support is scaffolding and doesn't work yet.
• Agent Context users need a Fivetran account, and query-only users leave the Account Role empty. An admin adds them under MCP Access > Users.

What can a Fivetran MCP server read and change?

Agent Context MCP lists two tools: read-only execute_sql and provide_feedback, which submits feedback on an answer. fivetran-mcp builds tools from Fivetran's API and defaults to the read scope. read/write adds writes, read/write/delete adds deletes, and DISALLOWED_ACTIONS blocks chosen endpoints inside that scope.

What it can read

• Agent Context MCP: execute_sql runs read-only SQL on your destination and uses the AGENTS schema to find context. An admin can set a read-only role on Snowflake or a read-only service account on BigQuery.
• fivetran-mcp: <resource>_read tools such as connections_read, plus list_endpoints and get_schema for discovery.
• fivetran-mcp resources include connections, destinations, groups, transformations, users and webhooks.
• Credential endpoints are reachable too. Fivetran's README says to deny system-keys:read, users:read:get_user_api_key and users:read:list_api_keys with DISALLOWED_ACTIONS.

What it can change

• Agent Context MCP: provide_feedback. Fivetran lists no tool that changes warehouse data.
• fivetran-mcp at read/write: connections_write covers sync_connection, resync_connection, create_connection and modify_connection, and transformations_write covers run_transformation.
• fivetran-mcp at read/write/delete: delete_connection, delete_destination and delete_group.
• Write and delete tools ask the model to confirm first, and the server doesn't enforce that. Scope and DISALLOWED_ACTIONS are what stop a call.

How do you set up a Fivetran MCP server in Claude, ChatGPT and Cursor?

For Agent Context MCP, an admin builds a Context Layer and adds you under MCP Access, then you connect your client. Fivetran documents Claude, ChatGPT and Cursor for it. fivetran-mcp is a local server, and its README documents Claude Desktop, Claude Code, Codex and Cursor, with no ChatGPT steps.

{
  "mcpServers": {
    "fivetran": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/fivetran/fivetran-mcp", "fivetran-mcp"],
      "env": {
        "FIVETRAN_API_KEY": "your-api-key",
        "FIVETRAN_API_SECRET": "your-api-secret",
        "FIVETRAN_SCOPE": "read",
        "DISALLOWED_ACTIONS": "system-keys:read,users:read:get_user_api_key,users:read:list_api_keys,users:write:create_user_api_key,users:write:rotate_user_api_key,users:delete:delete_user_api_keys"
      }
    }
  }
}

{
  "mcpServers": {
    "fivetran": {
      "command": "uvx",
      "args": ["--from", "git+https://github.com/fivetran/fivetran-mcp", "fivetran-mcp"],
      "env": {
        "FIVETRAN_API_KEY": "your-api-key",
        "FIVETRAN_API_SECRET": "your-api-secret",
        "FIVETRAN_SCOPE": "read",
        "DISALLOWED_ACTIONS": "system-keys:read,users:read:get_user_api_key,users:read:list_api_keys,users:write:create_user_api_key,users:write:rotate_user_api_key,users:delete:delete_user_api_keys"
      }
    }
  }
}

What can you ask a Fivetran MCP server?

Start with pipeline health, because fivetran-mcp reads it at its default scope. Its README's example questions ask which connections are failing and when one last synced. The warehouse prompt needs Agent Context MCP and a built Context Layer. The sync prompt needs FIVETRAN_SCOPE set to read/write, so save it for later.

• Reads

  Which connections are failing right now? List each one with its status and last successful sync.

  Copy

• Reads

  When did the Salesforce connection last sync, and did it succeed?

  Copy

• Reads

  Show the schema config for the HubSpot connection: which tables are enabled and which are excluded. Don't change anything.

  Copy

• Reads

  Using our context layer, what was paid search spend last month? Show the SQL and the tables you used.

  Copy

• Writes

  Trigger a sync on the Google Ads connection. Confirm the connection with me first.

  Copy

What are the limits of a Fivetran MCP server?

During Public Preview, Agent Context MCP allows 1,000 requests an hour and 100,000 a day per account, and Fivetran says those limits can change. Context Layer needs a North America destination on Snowflake, BigQuery or Databricks. fivetran-mcp uses your REST API limit, which is 20,000 requests an hour on Free and paid accounts.

• Trial accounts get 500 REST API requests an hour. Past a limit, the API returns 429 with a Retry-After header.
• Context Layer allows up to 50 connections in total across documentation sources, systems of record and analytics sources.
• Fivetran says Agent Context's tool interface is subject to change.
• Role-based query access isn't supported on Databricks yet, so the read-only role option is for Snowflake and BigQuery.
• fivetran-mcp in HTTP mode runs without DNS-rebinding protection unless you set allowed origins or hosts.

How much does a Fivetran MCP server cost?

fivetran-mcp is free under its MIT license. Fivetran hasn't published a price for Context Layer or Agent Context MCP, and its pricing page doesn't mention them. The plan underneath bills on monthly active rows (MAR), and the Free plan covers 500,000 a month. Your warehouse bills the execute_sql queries.

• Context Layer needs a Standard, Enterprise or Business Critical plan, so the Free plan doesn't qualify.
• Fivetran adds a $5 base charge to standard connections between 1 MAR and 1M MAR.
• Fivetran counts a row once a month, however often it syncs, so a manual sync adds MAR for rows it hasn't already counted that month.
• Historical re-syncs don't count towards paid MAR on the pricing launched on 1 March 2025. The rows still load into your warehouse, which does that work.
• run_transformation can add model runs. 5,000 a month are free on all plans, then $0.01 each up to 30,000.
• Agents connected directly to a Snowflake or Databricks warehouse use your own OpenAI API key for semantic search.
• Our Fivetran pricing guide covers MAR rates, and Fivetran implementation covers sync frequency and cost controls.

What are the alternatives to a Fivetran MCP server?

The main alternatives are a warehouse server for the data Fivetran loads, or Fivetran's REST API for scripted checks. A warehouse server answers reporting questions without touching Fivetran. Fivetran also documents connecting agents directly to the warehouse, and says that can raise token use. MCP vs API covers when a script fits better.

• Snowflake MCP: Query what Fivetran loaded, if Snowflake is your destination.
• BigQuery MCP: Query what Fivetran loaded, if BigQuery is your destination.
• dbt MCP Server: Lineage, tests and job runs for the models built on Fivetran data.
• Fivetran REST API: Scripted sync checks with no model in the loop.

Which skill goes with a Fivetran MCP server?

Use the free Agent write safety skill before you raise FIVETRAN_SCOPE, because a write scope can start syncs and change connections. Use the Marketing context kit to record your connection names, destinations and owners once, so the assistant has less to guess when you ask which pipeline feeds a report.

• Agent write safety: Sets approvals before an agent can start syncs or change connections.
• Marketing context kit: Holds your connection names and data definitions in one folder for every prompt.