the best analytics MCP servers, and the one honest question to ask of each

an analytics MCP server is a small program that exposes your analytics data to an AI assistant like claude or cursor through the model context protocol, so you can ask "where do people drop off?" in your editor and the AI can actually go read your numbers instead of guessing. several analytics tools now ship one: PostHog has an official MCP server, Umami has a community one, and smolanalytics has 47 tools and 13 prompts built in. they are all real and all useful. the honest question that separates them is not how many tools they expose, it is whether the number the AI reads back to you is computed or generated. smolanalytics is the one where the AI cannot make up the number, because every answer is the same deterministic report the dashboard renders, and a CI test fails the build if the two ever disagree. and because you bring your own model, the AI part is free.

what is an analytics MCP server, really?

the model context protocol is a standard, opened by anthropic, that lets an AI assistant call out to tools and data sources in a consistent way. an analytics MCP server is just an MCP server whose tools happen to be your analytics: "list top pages", "get funnel conversion", "show retention for last week". you connect it once to cursor or claude code or the claude desktop app, and from then on you can ask questions in plain english and the assistant answers from your real data instead of from whatever it remembers about analytics in general.

the reason this matters is that it moves analytics out of a separate tab and into the place you already work. you are debugging a signup flow in your editor, you ask "how many people finished signup this week?", and the answer arrives without you leaving. if you want the longer definition with the moving parts spelled out, i wrote a glossary entry for the analytics MCP server.

which analytics tools ship an MCP server?

more every month, which is the good news. here is a fair look at the ones people actually ask about, with only things that are true and widely known.

PostHog ships an official MCP server. PostHog is a large, mature product analytics platform, so the surface it can expose is broad: events, insights, feature flags, and more. if you already run PostHog and want your AI to reach into it, the official server is the natural choice, and it inherits PostHog's full feature set.

Umami has an MCP server in the community. Umami is a well-liked, privacy-friendly, open-source web analytics tool, and the MCP server lets an assistant query the web-analytics side of it. if you want lightweight, cookieless web stats and an AI that can read them, it is a reasonable pairing.

smolanalytics ships an MCP server with 47 tools and 13 prompts, covering both web and product analytics (visitors, referrers, funnels, retention, paths, cohorts) from one snippet. it is a single go binary, so there is no separate data warehouse to run beside it. the wedge is not the tool count, it is what the tools return, which is the next section.

i am deliberately not listing tools i cannot verify ship a real MCP server, and i am not going to invent flaws in the ones above. PostHog and Umami are good software. the differences below are about a design choice, not a quality gap.

which one can the AI not lie with?

this is the question i care about most, and it is the one thing genuinely different about smolanalytics, so let me be precise about it rather than hand-wave.

most AI-over-analytics setups work like this: the AI writes a query against your data, runs it, and reads back the result. that is powerful, but it means the number the AI shows you came from a query the AI generated in the moment. a generated query is a second, separate path from your dashboard, and the two can quietly disagree, so the assistant can hand you a confident, wrong number and you would not know.

smolanalytics does not let the AI write the query. the MCP tools call the exact same deterministic reports the dashboard renders. so the AI returns the real computed figure or nothing, never a figure it composed. and this is not just a claim in a readme: there is a CI test that asks the same question through the dashboard and through the AI and fails the build if the two answers differ. that is what "the AI cannot hallucinate the numbers" actually means here, mechanically, not as a slogan.

if your analytics MCP server generates queries, you get flexibility, and that is a real, legitimate trade. if it reads from fixed reports like smolanalytics does, you get a number you can act on without double-checking. neither is wrong. i built for the second because analytics you cannot trust is worse than no analytics.

does the AI part cost extra?

with smolanalytics, no, and this is worth being clear about because it is unusual. some AI-analytics products charge you for the AI, because they run the model on their side and pass the cost along. smolanalytics is bring-your-own-model: you connect it to the AI you already pay for or run (claude, cursor, a local model), so the assistant that answers your questions is the one you already have. the analytics tool does not add an AI bill on top.

that also means your questions and your data go to your model, not through some middle layer that reserves the right to train on them. it is a smaller, more honest arrangement, and it is the same reason the MCP overview page leads with "your model, your data."

so which should i pick?

pick by what you already run and what you need the AI to do, not by a leaderboard.

if you are deep in PostHog and want your assistant to reach its full feature set, use PostHog's official MCP server. if you want simple, privacy-friendly web analytics with an AI that can read it, Umami plus its community server is a clean fit. if you want one small binary that does both web and product analytics, that you can self-host free, where the AI provably cannot invent the number and does not cost extra because you bring your own model, that is the case for smolanalytics.

the honest limits, since i will not pretend otherwise: smolanalytics does not do session replay, feature flags, experiments, heatmaps, or surveys. if any of those is what you actually need, PostHog is a more complete product and you should use it. smolanalytics is the focused choice for "just tell me the true number when i ask," and it is deliberately not trying to be everything.

the honest close

this is one person, and it launched recently. there are no testimonials on the site because there are none yet, and none will be invented. the reason to try smolanalytics is not a logo wall. it is that the numbers are computed, not generated, and there is a test that proves the AI's answer matches the dashboard's.

it is open source under MIT, a single go binary with no kafka or clickhouse beside it. self-host free forever, or hosted from $9/mo if you would rather not run a server, with a 14-day trial and no card. you can run the demo right now with docker run -p 8080:8080 ghcr.io/arjun0606/smolanalytics, and the code is at github.com/Arjun0606/smolanalytics. if you want to see how the MCP side connects, start on the MCP page.

the best analytics MCP server is the one you can trust the answer from. that is the whole reason i built this one.