Sur cette page

How to connect Amplitude MCP to ChatGPT

Connect Amplitude MCP to ChatGPT so you can ask questions about your product data, such as funnels, cohorts, and experiments, without leaving the ChatGPT interface. Setup uses OAuth and takes about two minutes. Use this when you want conversational access to Amplitude data from ChatGPT.

Use ChatGPT if you prefer an AI assistant interface in the browser or desktop app. If you want Amplitude MCP inside a code editor instead, use the Cursor connector.

You need an active Amplitude account with project access and a ChatGPT account on an eligible plan with developer mode access.

How do I connect Amplitude MCP in ChatGPT?

For general MCP setup reference, refer to the OpenAI remote MCP documentation.ChatGPT developer mode supports full Model Context Protocol client access for read and write operations. OpenAI documents developer mode as a beta feature for eligible ChatGPT plans.
  1. Navigate to ChatGPT or open the ChatGPT desktop app.
  2. Go to Settings > Apps & Connectors > Browse Connectors.

  3. Select Amplitude, then select Connect to start the OAuth connection.

  4. Complete Amplitude OAuth authorization when prompted.

  5. For best results, create a ChatGPT project specifically for Amplitude and add this system prompt to its instructions:

    text
    When using Amplitude MCP, follow these rules then act quickly and autonomously:
    - Use tools to find answers: If you need info (events, properties, chart definitions, cohorts), use tools to discover it rather than asking the user. Trust the Amplitude MCP tools provide access to actual data behind charts, dashboards, and other entities. Always attempt using tools before saying they don't exist.
    - Try NOT to ask clarifying questions: Make your best judgment with information available but sparingly elicit clarification from users
    - Resolve ambiguity yourself: When multiple options exist (e.g., which project to use, which saved chart or event matches best, how to define a segment), choose the most reasonable option based on tool results and context. Search saved charts, metrics, and other data before creating something new.
    - When responding to requests that involve Amplitude objects (charts, dashboards, or any entity), don't stop at referencing IDs and metadata. Retrieve underlying data, run analysis based on it, then share specific metrics as part of your analysis.
    - Complete the request: Execute the workflow requested, proactively share relevant data when analyzing, don't stop partway to ask for confirmation, then provide data-backed, actionable, and concise answers.
    - Report what you did: After completing the task, briefly explain key assumptions or data used
    - Cite your sources: When referencing data from Amplitude, include the link as part of the markdown response (()[])
    
  6. Start asking questions about your Amplitude data.

How do I verify the connection works?

Ask ChatGPT: "What Amplitude projects can I access?" If you see a list of your projects, MCP is connected and authenticated.

Do I need developer mode enabled to use Amplitude MCP in ChatGPT?

Yes. ChatGPT developer mode is required for full MCP client access, including both read and write operations. OpenAI documents this as a beta feature available on eligible ChatGPT plans. Without developer mode, MCP tool calls aren't available.

How do I switch Amplitude MCP from the US server to EU residency?

The ChatGPT connector uses the standard Amplitude OAuth flow, which routes to the correct region based on your Amplitude account. If you need explicit EU routing, contact Amplitude support.

What permissions does the Amplitude MCP server have?

The MCP server inherits the Amplitude permissions of the user who authenticates. You can only read or modify projects you already have access to in Amplitude. Amplitude grants no additional access.

Scoping the system prompt to a project keeps the Amplitude instructions from affecting other ChatGPT conversations. It also lets you share the project configuration with your team so everyone uses the same query behavior.

Was this helpful?