Connect Amplitude to Claude in Ten Minutes, Then Ask It These Five Questions

You already have the answers. They are sitting in Amplitude, behind four clicks, a chart builder and a filter you set up in March and no longer remember the reason for.

The Amplitude MCP server lets Claude ask for them instead. The Amplitude MCP tools are around fifty functions that let an AI client search your project, query charts, build cohorts, read your event taxonomy, run and inspect experiments, and pull session replays, all inside the permissions your Amplitude login already has. Setup takes ten minutes. What you ask first matters more.

Disclosure: I work at Amplitude. Views are my own, and everything below comes from Amplitude’s public docs, fetched on 28/09/2026.

What are the Amplitude MCP tools?

Amplitude’s MCP server docs group them into families. The ones you will touch in week one:

• Discovery and context - finding things in your project, starting with get_amplitude_context
• Charts and queries - asking questions of your data with query_amplitude_data
• Taxonomy - reading and tidying your events with get_amp_taxonomy, manage_amp_events and manage_amp_properties
• Cohorts and users - building cohorts and looking up a user
• Experiments and flags - reading experiment results and flag state
• Session Replay - listing and watching replays

There are more: dashboards, notebooks, surveys, feedback, agent analytics and the data warehouse. The docs list every tool by name, and that page is the one to trust over this one.

Two details from the docs worth knowing before you start. It uses your existing Amplitude permissions, with separate “Use MCP (read)” and “Use MCP (write)” controls, so Claude cannot see anything you cannot. And it is not an ingestion endpoint. It reads and manages your analytics. It does not send your production events.

How do you connect Amplitude to Claude?

In the Claude app or claude.ai, per Amplitude’s Claude guide:

1. Settings, then Connectors, then Browse.
2. Search for Amplitude and pick the verified connector.
3. Complete the Amplitude OAuth prompt.

On a Team or Enterprise plan, an owner has to enable the connector in organisation settings first. That is usually the ten minutes.

In Claude Code, per the Claude Code guide, one line:

claude mcp add -t http -s user Amplitude "https://mcp.amplitude.com/mcp"

Swap in https://mcp.eu.amplitude.com/mcp for EU residency. Then type /mcp and follow the OAuth flow. The -s user flag makes it available in every session, which is convenient. Hold that thought.

Why fifty tools is a thing to think about

Every tool you connect costs the model context before it has read a word of your question. Tool names, descriptions, schemas. It adds up.

Amplitude’s docs are candid about this. ChatGPT gets a curated subset of about 25 tools “due to a token cap”, and there is a progressive discovery variant of the server URL (?discovery=progressive) so the client does not load the lot up front.

I learned the size of this problem the stupid way. The agents that run my side company run headless on a schedule. Last week I found every one of them had been failing before the model ran, because each call loaded every connector I had ever installed, and the combined tool definitions pushed it past the context limit. The jobs kept starting on time. They kept finishing on time. Nothing in between happened.

It was, technically, the fastest the system has ever been.

The fix was a flag that loads only the connectors a job actually needs. You have this somewhere too, I would bet on it: a user-scoped server added once, for one task, now riding along on every session. Connect Amplitude where you do analysis. Leave it out where you do not.

The five questions to ask it first

The instinct is to open with the big one. “Why did conversion drop?” Resist it. An agent will answer that question confidently whether or not your data can. Ask these first, in this order.

1. “What Amplitude projects can I access?” The docs’ own connection test. It also tells you which project Claude will default to, before it quietly answers from staging.

2. “List the events in this project, grouped by what they seem to describe, and flag any that look like duplicates.” This hits the taxonomy tools. Most projects that have lived a few years have Sign Up, signup_completed and user_signed_up, each loved by a different team. If you do not know which one is real, neither does Claude.

3. “For our main activation chart, tell me exactly which event, filters and segments it uses.” Before you trust a number, trust the definition. This is the question that finds the filter from March.

4. “Where do people drop off between sign-up and the first action that shows value?” Now you are querying. Because questions two and three came first, you can check which events it chose and push back when it guesses.

5. “Show me three session replays of people who dropped at that step.” A funnel tells you where. A replay tells you what it looked like from the other side of the screen. This is the step that turns a chart into a decision.

Notice the pattern. Three questions about the data, two about the product. Most people do it the other way round, and then spend a week defending a number built on the wrong event.

The part worth keeping

Connecting an agent to your analytics is ten minutes of work. Deciding what it should know first is the actual job.

Give it the smallest set of tools it needs. Make it prove it understands your definitions before it tells you anything about your users. Then let it go fast, because by then fast is safe.