Tinybird MCP Server

The Tinybird MCP server connects AI assistants to Tinybird, a platform for real-time analytics on large data volumes. Anyone collecting event data, logs, or product metrics in Tinybird can ask the assistant questions like "How many users were active in the last hour?" or "Which endpoints return the most errors?". The assistant calls published endpoints or writes a SQL query, runs it, and explains the result in the conversation. The server is operated by Tinybird itself and described in the official documentation as an access path for so-called analytics agents.

What is the Tinybird MCP server?

The server implements the Model Context Protocol (MCP), an open standard through which AI applications talk to tools and data sources. It runs hosted at https://mcp.tinybird.co and uses streamable HTTP. No local installation is needed. Clients that do not handle streamable HTTP themselves can, according to the documentation, be connected through the mcp-remote bridge. You need a Tinybird account with a workspace and a token whose permissions match the tools you want.

Which tools the server provides

The documentation lists six tools. With list_endpoints and call_endpoint the assistant finds published endpoints with their parameters and calls them; the result arrives as CSV by default, or as JSON or Markdown. list_datasources and list_service_datasources show the workspace's data sources and Tinybird's service data sources. With execute_query the assistant runs SQL through the Query API and can have the query plan shown, and text_to_sql translates a natural-language question into SQL. Which of these tools are usable depends on the token: a JWT or a resource-scoped token allows only selected endpoints or data sources, while an admin token opens all endpoints.

Setup and token

For Claude Code a single command with the HTTP transport is enough; in Cursor you enter the address in the MCP configuration. According to the documentation the token is appended to the address as the token parameter. If the workspace is not in the default region, the host parameter with the regional API address must be set as well, because a JWT carries no region information. For Claude Desktop the documentation has no dedicated instructions.

Security and limits

Because the token is part of the address, it can show up in configuration files, history data, or logs. Treat the complete address like a password and do not commit it to a repository. The documentation explicitly recommends using an admin token only for trusted internal tools. For an AI assistant a narrowly scoped token or JWT limited to individual endpoints or data sources is the better choice. Query results enter the language model's context and thereby reach the model provider; if your data contains personal information, check that beforehand. You should read an automatically generated SQL query before trusting its result, since it can consume compute time and therefore cost.

Costs

The documentation names no separate fee for the MCP server. Usage counts toward the respective Tinybird plan. According to the pricing page there is a free entry tier without a credit card and with limited resources, above it the paid Developer and SaaS plans on shared infrastructure, and Enterprise with a dedicated cluster. Current terms are on the official pricing page. The documentation makes no statement about the server's license.

Who is the Tinybird MCP server for?

It is worthwhile for teams that already use Tinybird for product or operations analytics and want to put questions to the data without detouring through a dashboard or SQL editor. It is especially useful when endpoints are already published: the assistant then uses vetted, parameterized queries instead of free-form SQL. Anyone not using Tinybird gains nothing, because the server is an access path to the platform and not standalone analytics.

A Tinybird account with a workspace, a suitably scoped token, and an MCP client with streamable HTTP support (or mcp-remote).

The Tinybird MCP server runs hosted by Tinybird. This guide walks through setup, token choice, and the first queries.

Step 1: Prerequisites

You need a Tinybird account with a workspace, a token, and an MCP client that supports streamable HTTP. Clients without that support can be connected through npx -y mcp-remote.

npx -y mcp-remote

Step 2: Choose a token

Create a narrowly scoped token or a JWT for the assistant that exposes only the endpoints or data sources it needs. According to the documentation an admin token is meant only for trusted internal tools.

Step 3: Set up in Claude Code

Run claude mcp add tinybird --scope user --transport http "https://mcp.tinybird.co?token=TINYBIRD_TOKEN" and replace the placeholder with your token. If your workspace is not in the default region, append &host= with the regional API address, for example https://api.eu-central-1.aws.tinybird.co.

claude mcp add tinybird --scope user --transport http "https://mcp.tinybird.co?token=TINYBIRD_TOKEN"

Step 4: Set up in Cursor

In ~/.cursor/mcp.json (global) or .cursor/mcp.json (project) add an entry tinybird under mcpServers with the url https://mcp.tinybird.co?token=TINYBIRD_TOKEN. Do not put the file containing the token into a repository.

Step 5: First queries

Let the assistant list the published endpoints and the data sources first. Then ask for a metric that an existing endpoint delivers, and for freely generated SQL check the query before you trust the result.

Tinybird token (static, resource-scoped, JWT, or admin token) as the token parameter in the address.

Call endpoints, list data sources, run SQL, and translate questions into SQL; the scope follows the token's scopes. A JWT restricts access to selected endpoints, an admin token opens all of them.

The server runs hosted by Tinybird. Query results enter the AI client context and therefore reach the model provider.

Token as part of the address (risk in configurations and logs), overly broad admin tokens, personal data reaching the model provider, and compute-heavy, freely generated SQL queries.