# Mixpanel MCP server: connect Claude and get product answers you can check

Canonical URL: https://www.martechdevs.com/blog/mixpanel-mcp-server

By [Jafrin Akter Alo](https://www.martechdevs.com/authors/jafrin-akter-alo)

Data Engineer and Co-founder, martechdevs

Published: 2026-05-21

Last updated: 2026-10-12

Mixpanel MCP lets an AI assistant investigate product data while you work. The useful outcome is a defensible answer: which customers reached value, where the journey broke and what deserves attention next. Connecting the server is only the first step. You still need the right project, a precise metric and a way to check the result.

## Understand what you are connecting

The server queries analytics and writes dashboards, Lexicon, cohorts and metrics using your existing project permissions. [Mixpanel's MCP documentation](https://docs.mixpanel.com/docs/mcp) lists the tools.

Start with one recurring question. For a product lead, that might be whether new workspaces reach their first successful integration. For growth, it might be which acquisition cohorts retain after their initial activation. A founder might need to distinguish more signups from better customer quality.

Write down the decision the answer should change. If every possible result produces the same action, improve the question before building an automated reporting workflow.

## Check access, then choose the regional endpoint

MCP defaults on for Free/Growth accounts created after 1 August 2026. Otherwise, an admin enables it under Settings → Org → Overview; changes can take 15 minutes. Free Query API limits can apply. See the [server guide](https://docs.mixpanel.com/docs/mcp).

Use the endpoint matching your project's data region:

- US: `https://mcp.mixpanel.com/mcp`
- EU: `https://mcp-eu.mixpanel.com/mcp`
- India: `https://mcp-in.mixpanel.com/mcp`

Have the administrator confirm which projects and actions the connecting identity needs. A growth investigation usually does not require permission to reorganise every team's definitions. Check the actual role and client controls; adding “read only” to a prompt does not remove account permissions.

Also decide which AI workspace may receive the results. [Mixpanel's AI policy](https://mixpanel.com/legal/ai-compliance/) describes its own AI processing arrangements. Do not assume those arrangements automatically govern a separate assistant account your company connects. Review that account's contract and configuration, especially before retrieving individual profiles or replay-derived information.

## Connect through Claude or Claude Code

In Claude's Settings → Connectors, add the regional URL and complete OAuth sign-in. Follow the [Mixpanel setup guide](https://docs.mixpanel.com/docs/mcp).

For Claude Code, use the remote HTTP syntax documented by [Anthropic](https://code.claude.com/docs/en/mcp). This example configures the US endpoint:

```bash
claude mcp add --transport http mixpanel https://mcp.mixpanel.com/mcp
```

Replace the URL for EU or India before running it. Inside Claude Code, open `/mcp`, select the connection and complete authentication. The command adds configuration; it does not establish that the account can answer your intended query. Anthropic's documentation explains connection status and reauthentication.

Do the first check in a new conversation:

> List the projects available through this connection. Show their IDs and names. Do not create or modify anything. Wait for me to select the production project before analysing data.

Select by ID as well as name. A project called “Production” can still belong to the wrong organisation, and a staging project can contain plausible-looking test data.

Then request one existing report you already understand. Ask the assistant to explain the metric, time range and filters before discussing its result. Compare the returned definition and numbers with the report in Mixpanel. This establishes a useful baseline without inventing a new analysis during the connection check.

If authentication succeeds but access fails, confirm the organisation setting, selected region and project membership. Capture the failing operation and error for the administrator. Repeating the same vague question is unlikely to reveal which access condition is missing.

## Give the assistant the definitions your team actually uses

[Business Context](https://docs.mixpanel.com/docs/business-context) supports organisation and project notes. Organisation admins manage company-level context; project owners and admins manage project context. Use the company layer for business vocabulary and the project layer for the actual measurement rules.

Here is an illustrative context note for a B2B product. It is an analysis specification, not a Mixpanel API payload:

```yaml
decision: identify where new workspaces fail to reach initial value
environment: production
entity: workspace
identifier: workspace_id
cohort_entry: first_workspace_created
success: first_integration_connected
success_window: 14 days after cohort entry
timezone: UTC
exclude: internal workspaces and approved test accounts
reconnections: do not count as first integration
required_breakdown: acquisition_route_at_workspace_creation
unknown_values: retain as a separate group
output: counts, rates, definitions, missing data and evidence links
```

The note needs an owner who can explain why these rules exist. “Activated user” is insufficient when the subscription belongs to a workspace. “Integration connected” is insufficient when one implementation fires on a button click and another fires after a successful backend response.

Ask the assistant to discover the available events and properties and map them to the specification. If the account model or required historical property does not exist, stop the analysis at that gap. Counting people and relabelling them as workspaces would produce an answer to a different question.

Put verified trigger definitions, exclusions and field meanings into your [tracking plan](https://www.martechdevs.com/blog/tracking-plan). Context should point to maintained definitions rather than becoming a second, contradictory dictionary.

## Treat Lexicon as part of the analysis

[Lexicon](https://docs.mixpanel.com/docs/data-governance/lexicon) holds event and property descriptions, owners and saved definitions. Its Verified Data feature is documented for Enterprise customers. Confirm availability before making it part of a rollout requirement.

A useful description of `integration_connected` explains the successful state transition, whether it is emitted by the backend, which workspace it belongs to and how retries are represented. “User connected an integration” leaves the important questions unanswered.

For missing descriptions, have the assistant prepare proposed text with evidence from instrumentation or an accountable event owner. Review each proposal before saving it. A fluent description inferred from a name can turn an uncertain event into an apparently authoritative definition.

Verified Mode flags unverified entities and prefers verified equivalents without excluding data. The MCP audit-log tool is a limited-customer beta. Check availability in the [MCP guide](https://docs.mixpanel.com/docs/mcp).

Maintain your own change record for approved edits: entity identifier, old description, proposed description, evidence, approver and time. That record is useful even when a vendor audit log is available because it captures the reasoning behind the change.

## Ask for the query definition before the conclusion

Mixpanel funnels have configurable counting methods, ordering, re-entry, conversion windows and property attribution. Those settings change the question being answered. [The funnel documentation](https://docs.mixpanel.com/docs/reports/funnels/funnels-advanced) is the reference when two apparently similar analyses disagree.

Use a prompt that makes the analysis reproducible:

> Investigate the agreed workspace activation metric for the July and August 2026 entry cohorts. Use UTC and allow each workspace its full 14-day observation window. First show the available event mapping, entity model, filters and proposed query settings. Identify missing prerequisites. After the definition is confirmed, return eligible counts, activated counts, rates, acquisition-route breakdowns and links or query details sufficient to reproduce the result. Separate observations from possible explanations. Make no project changes.

The observation window matters. A workspace created yesterday has not had the same opportunity to activate as one created three weeks ago. Either compare fully observed cohorts or clearly label an early-progress measure with the same age for every cohort.

Keep the cohort-entry dates separate from the dates on which success events can occur. An August entrant may activate in September. Restricting every event to August can accidentally cut off its allowed conversion window.

![A product question moves through project and event definitions, entity and observation-window selection, queries, reproduction and a reviewed decision.](https://www.martechdevs.com/assets/blog/mixpanel-mcp-server.svg)

## A worked example: an improving rate with no segment improvement

Consider fictional, fully observed workspace cohorts. Assume every eligible workspace has one stable identifier, its acquisition route is recorded at creation and both routes use the same activation definition.

In July, 800 self-serve workspaces produce 320 activations. Another 200 sales-assisted workspaces produce 160. The combined rate is 480 / 1,000, or **48%**.

In August, 600 self-serve workspaces produce 240 activations. Another 400 sales-assisted workspaces produce 320. The combined rate is 560 / 1,000, or **56%**.

The headline improved by eight percentage points. Self-serve activation stayed at 40%; sales-assisted activation stayed at 80%. The aggregate changed because the higher-converting route became a larger part of the cohort.

An answer saying “the onboarding redesign increased activation” would go beyond this evidence. A useful answer would identify the mix change and ask whether the extra assisted acquisition justified its cost. The product manager would still need evidence that the redesigned experience improved outcomes within a comparable population.

Ask the assistant to return the counts behind every rate, then calculate the rates independently. Check that segments reconcile to the eligible population. If 80 workspaces have no acquisition route, retain them as an unknown group; excluding them silently can change the headline.

This is an original arithmetic example, not a Mixpanel account result or an experiment. It shows why a convincing explanation must survive a denominator check. The [activation metrics guide](https://www.martechdevs.com/blog/saas-activation-metrics) develops the measurement design further.

## Turn an investigation into a usable team output

Once the result checks out, request a compact decision note: the question, metric definition, evidence, uncertainty, proposed action and owner. Link the underlying report or preserve the query configuration so another person can inspect it.

For example, “investigate assisted-acquisition cost before scaling that route” is actionable. “Improve onboarding and optimise conversion” does not specify what the team should do or which evidence would justify it.

If a dashboard or cohort is useful, review its proposed contents and intended users before creation. Use a clearly named draft location. Inspect the saved object afterwards, including filters, sharing and description. Saving a report should preserve the definition you approved, not merely reproduce the headline number.

Keep repeated investigations bounded. Begin with one metric and one comparison, then expand only when a finding calls for it. This controls review effort and prevents a broad prompt from producing dozens of unrelated charts nobody will maintain.

## Where the Mixpanel skill pack and Mixpanel Agent fit

The official [Mixpanel AI plugin repository](https://github.com/mixpanel/ai-plugins) provides reusable workflows for report analysis, metric investigation, dashboards, Lexicon and tracking work. Its current plugin is named `mixpanel`; older region-specific plugins have been retired. Region is configured during setup.

For Claude Code, the documented installation is:

```bash
claude plugin marketplace add mixpanel/ai-plugins
claude plugin install mixpanel
```

Then run `/mixpanel:install` inside Claude Code and choose the engine and region. The plugin supports MCP or the Mixpanel Headless SDK. You do not need the plugin simply to use a working MCP connection.

A skill supplies a workflow. Inspect its proposed actions and the permissions of the connected account before using it for cleanup or creation. Keep your metric definitions authoritative even when a packaged workflow suggests a different default.

[Mixpanel Agent](https://docs.mixpanel.com/docs/mixpanel-agent) is the assistant inside Mixpanel, with access to product analysis and content-creation capabilities. Use it when the work naturally lives alongside an existing report. An external MCP client is useful when the investigation also needs approved context from a repository or another connected system. Choose based on the job and the data you are prepared to expose to that client.

## A practical rollout checklist

- [ ] One recurring decision and an accountable reviewer are selected.
- [ ] The connecting identity has the intended project access.
- [ ] The regional endpoint and AI workspace are approved for the data involved.
- [ ] A known report reconciles through the connection.
- [ ] Entity, event triggers, exclusions and observation window are explicit.
- [ ] Missing properties and unverified definitions remain visible.
- [ ] Counts, denominators and segment totals are checked.
- [ ] Proposed content changes have a review and change-record process.
- [ ] Saved outputs preserve their definitions and evidence.
- [ ] The pilot records useful decisions and review effort, not just prompt volume.

If the blocker is inconsistent events or an unclear account model, fix that foundation before automating more questions. Our [product analytics work](https://www.martechdevs.com/product-analytics) covers tracking definitions, implementation and reconciliation so a conversational answer can be traced back to a measurement your team trusts.

## Frequently asked questions

### Is the Mixpanel MCP connection read only?

Do not treat it as read only. Plan the connecting role and client controls around the operations you intend to permit. A prompt requesting analysis is useful task direction, but it does not change the credentials behind the connection.

### Which endpoint should an EU project use?

Use `https://mcp-eu.mixpanel.com/mcp`. Check the project's actual region with its administrator instead of choosing an endpoint based on your own location.

### Can Claude fix missing event descriptions?

It can help draft them. Have the event owner or implementation evidence establish what actually fires, which properties are sent and where the definition applies. Review that evidence before saving a generated description.

### Does Verified Mode make every answer correct?

No. Human review of definitions does not establish that a chosen population, conversion window or causal explanation is appropriate. Keep checking the analysis even when its inputs carry a verification badge.

### Should we replace our weekly dashboard with MCP?

Keep the stable view people rely on until the conversational workflow reproduces it and improves a real decision. MCP is particularly useful for follow-up investigation. A shared, maintained report can still be the clearest reference for a recurring meeting.

## Sources

1. [Mixpanel MCP server](https://docs.mixpanel.com/docs/mcp)

2. [Mixpanel AI processing policy](https://mixpanel.com/legal/ai-compliance/)

3. [Claude Code MCP connections](https://code.claude.com/docs/en/mcp)

4. [Mixpanel Business Context](https://docs.mixpanel.com/docs/business-context)

5. [Mixpanel Lexicon](https://docs.mixpanel.com/docs/data-governance/lexicon)

6. [Mixpanel funnel settings](https://docs.mixpanel.com/docs/reports/funnels/funnels-advanced)

7. [Mixpanel AI plugins](https://github.com/mixpanel/ai-plugins)

8. [Mixpanel Agent](https://docs.mixpanel.com/docs/mixpanel-agent)