Product Analytics MCP Servers: 3 Compared and What They Automate

Most people never open their analytics tool. They ask someone who does. Product analytics MCP servers let anyone ask the question in plain English and get the chart.

I compared the three most used. PostHog, the open-source all-in-one. Mixpanel and Amplitude, the two classic event analytics platforms.

All three ship official remote servers. All three can also change things: dashboards, cohorts, tracking plans, experiments and feature flags.

That last one matters. A wrong answer about retention is annoying. A feature flag rolled out to every user by mistake is an incident.

This guide covers what each server can do, which analytics work is worth automating, and how to keep an agent away from the switches that change your product.

Every endpoint and tool name below comes from the vendor's own documentation, checked in October 2026.

What Is a Product Analytics MCP Server?

A product analytics MCP server exposes your analytics project as tools an AI model can call. Run queries, read dashboards, look up events and properties, inspect cohorts, experiments and flags.

New to the protocol? Start with what the Model Context Protocol is. The short version: MCP is a standard way for a model to discover and call external tools.

The value is speed. "How did the new onboarding change activation?" becomes one sentence instead of a ticket to the data team and a two-day wait.

The 3 Product Analytics MCP Servers Compared

Here is the side-by-side. Endpoints are quoted from each vendor's documentation, not from a directory listing.

1. PostHog MCP Server

PostHog's server is at https://mcp.posthog.com/mcp. It routes you to the US or EU region based on the account you sign in with, so there is one URL for both.

The quickest setup is the wizard, which configures Claude Code, Cursor, VS Code, Codex, Zed and others:

npx @posthog/wizard mcp add

OAuth is recommended. For clients without it, create a personal API key with the "MCP Server" preset, which scopes it to one project, and send it as a bearer token.

PostHog is the biggest server here, with hundreds of tools across product analytics, SQL, error tracking, session replay, feature flags, experiments, surveys and more.

Loading all of that would flood your context, so PostHog defaults most clients to CLI mode. The model gets a single exec tool and uses it to search, inspect and call PostHog tools on demand. Cursor defaults to tools mode, which loads every schema.

I explain why that matters in MCP context bloat.

The safety switches are the best of the three:

• ?readonly=true, or the x-posthog-read-only header, removes every tool that creates, updates or deletes
• ?features=insights,dashboards exposes only those feature areas, and ?tools= picks tools by exact name
• Pin the project with project_id, which also removes the switch-project tools

Connecting and calling tools is free. Some tools use an LLM internally and are billed as PostHog AI spend. The rate limit is 480 requests per minute and 4,800 per hour per credential.

Setup details are on our PostHog MCP server page.

2. Mixpanel MCP Server

Mixpanel has one endpoint per data region:

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

Use the one for the region your project lives in. Auth is OAuth with PKCE, with service accounts in beta for scripts and CI.

There is one step people miss. For most orgs, an admin has to switch MCP on in Settings, Org, Overview. Only new Free and Growth accounts created after 1 August 2026 have it on by default.

The core tools are Run-Query, which runs insights, funnels, flows and retention, and Get-Query-Schema. There are also tools for dashboards, cohorts, Lexicon events and properties, metrics, experiments and feature flags.

Many of those write: create and delete dashboards, bulk-edit events and properties, create and update cohorts, experiments and feature flags.

Mixpanel's strongest guardrail is visibility. Every change made through MCP is recorded in the audit log with MCP as the origin. Reads and failed calls are not logged.

Know the limits. The rate limit is 600 MCP requests per hour per user. Run-Query cannot filter by cohort or use Saved Metrics. And experiments and flags created over MCP cannot set audience targeting, so finish those in the UI.

3. Amplitude MCP Server

Amplitude's server is at https://mcp.amplitude.com/mcp, with https://mcp.eu.amplitude.com/mcp for EU data residency. Auth is OAuth 2.0, and any existing Amplitude customer can use it.

It has 45 tools: charts, dashboards, notebooks, cohorts, session replays, experiments, feature flags, the tracking plan, and more.

Add ?discovery=progressive to the URL and the model starts with a small set of discovery tools and loads the rest by category. That keeps context small.

Amplitude's safety model is permission-based, and it is the most thorough of the three:

• "Use MCP (read)" and "Use MCP (write)" are separate permissions per project and role. A role with read only gets a read-only server
• Admins can block MCP for the whole org under Settings, Content Access, MCP
• Deletes take two calls. The first returns a summary, and the client must call again with confirmed set to true
• Cohort syncs preview the export and ask for confirmation before sending users to Braze, Iterable, S3 or a webhook

One thing to know: this is not an ingestion path. manage_amp_events changes event definitions in the tracking plan. It does not send event data.

Setup details are on our Amplitude MCP server page.

Before you connect any of these to your real project, see what tools it returns. Test any MCP server free →

Analytics Work You Can Automate With MCP Servers

The best uses are questions someone would otherwise wait days to get answered.

Ad-Hoc Questions

The everyday win. Anyone on the team asks, and the agent builds the query.

Compare 7-day retention for users who signed up before and after
the new onboarding shipped on 1 October. Show the query you ran
and the event names you used.

Always ask for the query and event names. Analytics agents are confidently wrong when two events have similar names.

Weekly Metric Summaries

"What moved this week and why?" Read the main dashboard, compare to last week, and flag anything unusual. Read-only, so a good first job.

Experiment Readouts

All three can read experiment results. Ask for the result, the sample size and whether it is significant. Shipping the winner stays a human decision.

Tracking Plan Cleanup

Find events with no volume, duplicate properties and missing descriptions. Mixpanel and Amplitude can bulk-edit Lexicon and the tracking plan. Review every batch, because renaming an event can break dashboards.

Errors to Product Impact

PostHog has error tracking in the same server, so an agent can ask "how many users hit this error, and did they churn?" Pair it with an observability server for the stack-trace side.

What Not to Automate

• Feature flag rollouts. PostHog even has a "roll out to everyone" tool. Keep it away from agents
• Shipping or ending experiments. One call can end a test early
• Cohort syncs to ad and email tools. That sends user lists to third parties
• Deleting events or properties. Custom events delete permanently on Amplitude
• Bulk tracking-plan edits without a review

How to Pick a Product Analytics MCP Server

You use the server for the tool your team already has. Within that, here is what to look for:

1. Want a one-line read-only switch? PostHog, with ?readonly=true
2. Want access managed by role? Amplitude, with separate MCP read and write permissions
3. Need an audit trail of agent changes? Mixpanel logs every MCP write with MCP as the origin
4. Care about context size? PostHog's CLI mode or Amplitude's progressive discovery
5. Need EU or India data residency? Mixpanel has both endpoints, Amplitude has EU, and PostHog routes by account

Comparing other categories? See data warehouse MCP servers and SEO MCP servers.

The Feature-Flag Problem With Analytics MCP Servers

Analytics feels read-only. It is not, because these platforms also run your feature flags and experiments.

A flag change is a production change. It takes effect for real users in seconds, with no deploy and no code review.

An agent asked to "clean up old flags" can disable one that is still in use. An agent asked to "ship the winner" can roll a variant out to everyone.

Then there is the data itself. Person profiles, session replays and event properties can contain personal data, and all of it goes into the model's context when the agent reads it.

Four things that help:

• Read-only by default. PostHog ?readonly=true, or Amplitude's "Use MCP (read)" without write
• Filter out flags and experiments for analysis agents, with PostHog's ?features= or ?tools=
• Pin one project so an agent cannot wander into production from staging
• Watch the audit log. Mixpanel tags every MCP change, so review it weekly

The OWASP MCP Top 10 covers this under excessive permissions. Running your own analytics MCP server? Scan your MCP server →

How MCP Playground Can Help

Before connecting an analytics MCP server to your real project, see what it exposes.

The MCP server tester lists every tool with its schema. Connect PostHog with and without ?readonly=true and compare the two tool lists.

MCP Agent Studio runs real prompts with full JSON for each call. That shows you which events the model actually queried, which is how you catch a confident answer built on the wrong event.

You can also compare models on the same question. Some ask which event you mean. Some guess.

Frequently Asked Questions

Do PostHog, Mixpanel and Amplitude have official MCP servers?+

Yes. PostHog runs mcp.posthog.com/mcp and routes to the US or EU region by account. Mixpanel runs mcp.mixpanel.com/mcp with EU and India endpoints. Amplitude runs mcp.amplitude.com/mcp, with mcp.eu.amplitude.com/mcp for EU data residency.

How do I make the PostHog MCP server read-only?+

Add ?readonly=true to the URL or send the x-posthog-read-only header. Tools that create, update or delete data are removed. You can also limit tools with ?features= or ?tools=, and pin one project with project_id.

Why can't I connect to the Mixpanel MCP server?+

For most organizations an admin has to enable MCP in Settings, Org, Overview, and changes can take up to 15 minutes. Only new Free and Growth accounts created after 1 August 2026 have it on by default. Also check you are using the endpoint for your data region.

Can an analytics MCP server change feature flags?+

Yes. PostHog, Mixpanel and Amplitude all expose tools that create or update feature flags and experiments. Keep analysis agents read-only, or filter those tools out, because a flag change affects real users immediately.

Can I send events through the Amplitude MCP server?+

No. Amplitude says its MCP server is not an ingestion endpoint. Its event tools change definitions in the tracking plan. Send event data through an Amplitude SDK or the HTTP V2 API.

Conclusion

PostHog has the best switches, with read-only mode, feature filters and CLI mode. Amplitude has the most thorough permissions. Mixpanel has the clearest audit trail.

Start with read-only questions and weekly summaries, and ask the agent to show its query every time. Keep feature flags, experiments and cohort syncs with a human.

Before you connect one to your project, list its tools and watch a few calls run. Test any MCP server free →