Amplitude MCP server: how founders with a coding agent set it up in Claude Code, Codex and Cursor
Amplitude runs a hosted MCP server at https://mcp.amplitude.com/mcp (EU data: https://mcp.eu.amplitude.com/mcp). You sign in through OAuth, with no API key. In Claude Code run claude mcp add --transport http --scope user Amplitude https://mcp.amplitude.com/mcp, then /mcp to sign in. The agent then reads and builds charts, cohorts, experiments and flags, and reads session replays up to 30 days old, with your own Amplitude role.
This page is for founders who build with Claude Code, Codex or Cursor and already send events to Amplitude. We sent the server requests from our build machine on November 15, 2026, without an Amplitude login, to see what a coding agent meets before sign-in; the tool list and permissions below come from Amplitude's docs.
Jump to: setup · what we saw · the tools · who can use it · session replays · limits · when you need more · FAQ
How do you add Amplitude's MCP server to a coding agent?
| Client | Command or config | Then |
|---|---|---|
| Claude Code | claude mcp add --transport http --scope user Amplitude https://mcp.amplitude.com/mcp | Start claude, type /mcp, pick Amplitude, sign in. --scope project writes a shared .mcp.json instead. |
| Codex | codex mcp add amplitude --url https://mcp.amplitude.com/mcp | Sign in when asked; codex mcp login amplitude signs in again. The Codex IDE extension reads the same config.toml. |
| Cursor | {"mcpServers": {"Amplitude": {"url": "https://mcp.amplitude.com/mcp"}}} in .cursor/mcp.json (project) or ~/.cursor/mcp.json (global), or Amplitude's one-click install link | Cursor Settings, Tools & Integrations, then Connect or Needs login next to Amplitude. |
| Claude (desktop and claude.ai) | A custom connector with the same URL | Amplitude has a separate guide for it. |
Use the EU URL only when your Amplitude data lives in the EU region: Amplitude's docs say sign-in fails when the URL does not match your data region. Sign in to one Amplitude organization during the OAuth step. For a large tool list, Amplitude documents ?discovery=progressive on the URL, which has the agent load tools by category through list_tool_categories, get_category_tools and describe_tool.
What did we see when we tried it without signing in?
On November 15, 2026 we sent the server one MCP initialize request with no credentials, read its OAuth metadata, and added it to Claude Code 2.1.322 in an empty test folder with project scope. What came back:
| Check | Result | What it means for your agent |
|---|---|---|
initialize with no token | HTTP 401, "Missing authorization header", and a WWW-Authenticate: Bearer realm="OAuth" header that points to /.well-known/oauth-protected-resource/mcp | The client is told where to sign in, This is the standard pointer an MCP client follows to start its browser sign-in; Amplitude's docs describe that sign-in for Claude Code, Codex and Cursor, and we did not complete it. |
| Protected-resource metadata | Scopes all:read and all:write; authorization server https://mcp.amplitude.com | Access is all-or-nothing at the token level; Amplitude narrows it with your role and project permissions (below). |
| Authorization server metadata | A client registration endpoint (/register), PKCE with S256, scopes all:read, all:write and offline_access | A client can register itself and ask for a refresh token, so a sign-in can outlast one session. We did not complete a sign-in to confirm it. |
| EU server, same request | HTTP 401 | Same behaviour on mcp.eu.amplitude.com. |
claude mcp add ... --scope project, then claude mcp list | Amplitude: https://mcp.amplitude.com/mcp (HTTP) - ⏸ Pending approval (run `claude` to approve) | A project-scope server waits for each person to approve it in a session before it connects or asks for sign-in. |
We did not sign in, so we did not call any tool. If the server shows anything other than connected after sign-in, how to check if an MCP server is running reads each status line.
What tools does the Amplitude MCP server give the agent?
Amplitude's docs group the tools like this (ChatGPT gets a single search tool in place of the first two tools in the first row):
| Group | Tools | What they do |
|---|---|---|
| Search and context | search_amp_entities, search_amp_data_taxonomy, get_from_url, get_amplitude_context, manage_amp_context | Find charts, dashboards, cohorts, experiments and events by description; open whatever an Amplitude link points to; read or replace the project's AI context text. |
| Charts | query_amplitude_data, get_amplitude_charts, render_amplitude_chart, use_amplitude_chart_monitors | Ad-hoc segmentation, funnel and retention queries; read saved charts; render a chart with an edit link; set alerts. |
| Dashboards, notebooks, sharing | use_amp_dashboards, use_amp_notebooks, use_amp_comments, share_amp_entities, manage_amp_entities | Read, create and edit dashboards and notebooks, comments and sharing. |
| Cohorts and users | use_amplitude_cohorts, get_amp_user_data | Build and sync cohorts (a sync asks for confirmation); read a user's profile and activity. |
| Experiments, flags, metrics | use_amp_experiments, use_amp_flags, use_amplitude_metrics | Read, analyze, create and update experiments and feature flags; manage metric definitions and goals. |
| Tracking plan | get_amp_taxonomy, manage_amp_events, manage_amp_properties, manage_amp_branches, get_transformations, get_group_types | Read and edit event and property definitions and tracking-plan branches; deletes need a second confirmation. |
| Session Replay | get_amp_session_replay_info | Search, list and inspect replays (see below). |
| Feedback, AI Visibility, agents, warehouse | use_amplitude_ai_feedback, use_amp_ai_visibility, get_agent_results, use_amplitude_data_connections and others | Feedback themes, AI Visibility data, agent results, data connections and exports (some US-only). |
| Fallback | run_amp | Runs amp CLI commands against Amplitude's Developer API when no tool covers the job (guides and surveys go through it); deletes need allowDestructive: true. |
Who can use it, and what can it change?
- Amplitude's docs say any existing Amplitude customer can use the server from an MCP-capable client. Service accounts, for agents that run with no person signed in, come with Enterprise and are an add-on for Plus.
- Roles: the server acts as you and reaches only the projects you can already open. Two permissions, Use MCP (read) and Use MCP (write), are set per project: Viewer gets read; Member, Manager and Admin get read and write. Tracking-plan edits also need Update Tracking Plan.
- Admin switch: an Amplitude admin can allow or block MCP for the whole organization in Settings, Content Access, Model Context Protocol (MCP), and a user can't override it.
- Price: Amplitude's Free plan is $0 with 10,000 session replays a month and 30 days of replay retention. Plus starts at $0 and bills on events past the first 2 million a month. Growth has a custom price (Amplitude's pricing page, read November 15, 2026; details on Amplitude pricing).
Can your agent read Amplitude session replays?
Yes, as data. get_amp_session_replay_info searches replays up to 30 days old by user properties or events, lists them by time range, and returns a recording's interaction events and timeline. The agent reads that timeline as text; it does not watch video. Amplitude also publishes two skills for it: debug-replay, which turns a bug report into numbered repro steps from the timeline, and replay-ux-audit, which reviews several replays of one flow and ranks the friction. In Claude Code they come with Amplitude's plugin:
/plugin marketplace add amplitude/mcp-marketplace
/plugin install amplitude@amplitude
Amplitude's docs say the skills switch on by themselves when your request matches them. What they leave to you or the agent: which replays to open, and which change on your site to make.
Limits to plan around
- It does not send events. The server edits definitions; production event data still goes through an Amplitude SDK or the HTTP V2 API. If you have no Amplitude tracking yet, the agent has nothing to query.
- Region has to match. US URL for US data, EU URL for EU data, or sign-in fails.
- One organization per sign-in.
- Write access is real. A Member can let the agent create dashboards, flags and tracking-plan entries. Give the agent's user Viewer on projects where you only want reads.
- Rate limits: Amplitude's MCP docs state none; we sent too few requests to find one.
When you need more than the Amplitude MCP
Amplitude's server covers far more than UserProof's does: charts, experiments, flags, cohorts, the tracking plan and replays in one place, with your roles carried through. That is where it wins, and if your team already lives in Amplitude it is the one to connect. Its docs describe data access and replay skills for repro steps and friction maps; they do not describe a ranked fix that names the file and line of your code. UserProof does that other job. Its MCP server runs locally from a plugin and reads the visits it recorded. It hands the agent one fix at a time with the file and line, ranked by your signup or checkout goal, then checks whether the shipped fix worked. In Claude Code:
claude plugin marketplace add GetUserProof/userproof && claude plugin install userproof@userproof && claude "/userproof:setup"
UserProof is free for 3 sites and 5,000 recorded visits a month; Pro is $29 a month for 50,000 and Scale $99 for 250,000 (pricing). Compare the two on UserProof vs Amplitude, see Amplitude alternatives, or which session replay tools have an MCP server, each endpoint checked. For other servers worth wiring in, see the best MCP servers for coding agents, the Datadog MCP server for logs, traces and RUM replays, and the MCP config generator.
Sources, read November 15, 2026
- Our checks of
https://mcp.amplitude.com/mcpandhttps://mcp.eu.amplitude.com/mcp(initialize with no token, OAuth metadata) and our run ofclaude mcp addandclaude mcp list, Claude Code 2.1.322, November 15, 2026 - Amplitude, Amplitude MCP Server docs (URLs, OAuth, plans, permissions, the tool list, client setup, Session Replay skills; page last updated November 12, 2026)
- Amplitude, How to connect Amplitude MCP to Cursor
- Amplitude, Amplitude MCP product page (no production event data through MCP)
- Amplitude, pricing (Free, Plus and Growth replays), via our daily pricing ledger
- UserProof plans and setup: getuserproof.com/pricing, set up UserProof in Claude Code
FAQ
Does Amplitude have an MCP server?
Yes. Amplitude runs a hosted MCP server at https://mcp.amplitude.com/mcp, and https://mcp.eu.amplitude.com/mcp for data stored in the EU. You sign in with your Amplitude account through OAuth; there is no API key to copy. Amplitude's docs say any existing Amplitude customer can use it from an MCP-capable client.
How do I add the Amplitude MCP server to Claude Code?
Run claude mcp add --transport http --scope user Amplitude https://mcp.amplitude.com/mcp, start claude, type /mcp, pick Amplitude and finish signing in from your browser. Use --scope project instead to write a .mcp.json file your team shares; Claude Code then asks each person to approve the server before it connects.
Is the Amplitude MCP server free?
Amplitude's product page says the MCP server is available on every plan, including the free Starter plan, and Amplitude's Free plan costs $0 with 10,000 session replays a month (pricing page, read November 15, 2026). Service accounts for agents that run without a person signed in come with Enterprise and are an add-on for Plus. The AI app you connect may have its own cost.
Can Claude watch Amplitude session replays through MCP?
It can read them as data, not video. The get_amp_session_replay_info tool searches replays up to 30 days old by user properties or events and returns a recording's interaction events and timeline. Amplitude also ships two skills, debug-replay and replay-ux-audit, that turn those timelines into repro steps and a friction map.
Can the Amplitude MCP server change my data?
It can change definitions, not send events. With the Use MCP (write) permission, which Member, Manager and Admin roles carry, the agent can create dashboards, cohorts, experiments, flags and tracking-plan entries; tracking-plan edits also need the Update Tracking Plan permission; deletes ask for a second confirmation. It does not send production event data; that still goes through an Amplitude SDK or the HTTP V2 API.