smolanalytics

self-hosting posthog means clickhouse, kafka, redis and a postgres. this is one binary.

If you want product analytics you can actually run yourself, the options are bad. PostHog is a stack: ClickHouse, Kafka, Redis, and a Postgres to babysit. Plausible and Umami install easily but stop at web analytics, so no funnels, no retention, no cohorts.

smolanalytics is the middle that didn't exist. One Go binary, one data file, no external database. It does web analytics (visitors, referrers, UTM, devices) and product analytics (funnels, retention, paths, cohorts) from the same events, plus feature flags, A/B testing, click heatmaps, in-product surveys, a session inspector, and deploy-impact. Cookieless mode means no consent banner. MIT, free forever, and your data never leaves your box.

It also answers in plain English, from the dashboard or from your editor over MCP, using your own model, so the AI part costs nothing.

Try it in 30 seconds

docker run -p 8080:8080 ghcr.io/arjun0606/smolanalytics demo

# install script (macOS / Linux)
curl -fsSL https://raw.githubusercontent.com/Arjun0606/smolanalytics/main/install.sh | sh
smolanalytics demo

# or with Go
go run github.com/Arjun0606/smolanalytics/cmd/smolanalytics@latest demo

Open localhost:8080: a fully populated dashboard, a "what to fix" verdict up top, and an ask bar with your real events and pages as one-click chips. Nothing to configure. Prefer not to install? The live demo is the real product on demo data, running right now.

Ask your analytics where you write code

This is the point of the whole tool. smolanalytics is an MCP server, so your coding agent queries your real analytics without you leaving the editor. It has your codebase, your tracking plan, and smolanalytics over MCP, so it answers in your terms: ask "what's the MAU for the PQR page" and it knows PQR is the /pqr route from your code. Your model does the reasoning, so there are no API keys and nothing metered.

smolanalytics connect          # wires it into every coding assistant you have installed

you ▸ how's activation, and is pro converting better than free?
ai  ▸ Activation is 62% (657 of 1,051 signups reach "activate").
      Pro converts 2.4× better end-to-end: 45% signup→checkout vs 19% on free.
      The leak is activate→checkout on free (only 31% continue). Want the paths after activate?

Your model gets 79 tools and 14 built-in prompts, and it runs the whole product, not just queries: ask reports, roll out a flag, read the A/B result, create a cohort, set an alert, verify instrumentation. Anything it creates shows up on the dashboard instantly.

// stdio (local, reads your data file directly)
{ "mcpServers": { "smolanalytics": { "command": "smolanalytics", "args": ["mcp"] } } }

// HTTP (point at a running instance, local or remote)
{ "mcpServers": { "smolanalytics": { "url": "http://localhost:8080/mcp" } } }

curl -X POST http://localhost:8080/mcp -H "Authorization: Bearer YOUR_KEY" \
  -H "Content-Type: application/json" -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list","params":{}}'

There is also a built-in dashboard ask bar (zero setup, no code lookup) for quick data questions like "visitors to /pricing" or "where do people drop off?".

The whole toolkit, one binary

The usual advice is "run Plausible for web, something heavier for product, and separate tools for flags, experiments, and surveys." smolanalytics is all of it, computed from one append-only event log, and every surface is askable in your editor.

Full tool + prompt reference: docs/prompts.md · the plain-GET stats API: docs/api.md.

Provably correct, or it fails CI

The agreement test asserts, on every build, that the MCP answer equals the /v1 HTTP API answer byte-for-byte for the same question across funnel, retention, trends, web_overview, paths, heatmap, flag_impact, survey_results, list_sessions, and more. The dashboard renders from those same reports, so it cannot drift either. There is no second query path to disagree with. This is the one thing an AI-answer layer built on generated SQL structurally cannot promise.

Send events (web, mobile, server)

One snippet autocaptures pageviews + clicks. Add track() for the moments you care about.

<script src="https://YOUR_HOST/sdk.js"></script>
<script>
  smolanalytics.init("YOUR_WRITE_KEY", { host: "https://YOUR_HOST" });
  smolanalytics.track("signup", { plan: "pro" });   // optional, for funnels
  smolanalytics.identify("user_123");                // on login
</script>

Ingestion is one endpoint, so anything with an HTTP client works (curl -XPOST $HOST/v1/events ...). And there are published native SDKs with an offline-safe queue, batching, sessions, and lifecycle events:

Framework guides (2 minutes each): Next.js · React · Vue · Backend · Mobile. Or paste one line into Cursor / Claude Code and let it instrument the app: docs/agents.md.

Which deploy moved the metric?

Every other tool shows you the graph dropped. It cannot tell you which ship dropped it, because it does not have your commits. Record a marker in CI (smolanalytics deploy, one line) and ask your editor "did my last deploy move signups?" and you get a before/after read that leads with any regression, computed from the same reports (correlation, not proof, and the copy says so).

How it compares

Three things they structurally cannot copy: the AI is yours (they meter theirs); answers come from exact reports with CI proving they match the dashboard; your data never leaves your box.

Run it in production

docker run -p 8080:8080 -v $PWD/data:/data \
  -e SMOLANALYTICS_WRITE_KEY=$(openssl rand -hex 16) \
  -e SMOLANALYTICS_PASSWORD=$(openssl rand -hex 12) \
  ghcr.io/arjun0606/smolanalytics

One static binary, no cgo, no cluster. It binds 127.0.0.1 by default and refuses to serve real data unauthenticated on a public interface. Two keys: a public write key (ingest only, ships in your HTML) and a secret read key (reports, export, MCP). Scale to billions of events on flat RAM with an optional S3 / R2 / Tigris cold tier, keep a SMOLANALYTICS_RETAIN_DAYS window, and cron smolanalytics brief for a morning digest. Full config, backups, and the storage design: Deploy guide · STABILITY.md.

Own it, forever

• Private by architecture. No third party, no cookies by default, a cookieless mode that needs no consent banner, and GDPR erasure in one call (DELETE /v1/users/{id}/data). The answer to "who can see this data?" is: you.
• MIT, no CLA, no rug-pull. There is no license to revoke. Fork it the day you stop liking us.
• No lock-in. GET /v1/export hands you everything as CSV or JSONL, and the JSONL round-trips back into /v1/events. Import from PostHog, Mixpanel, Amplitude, Umami, CSV, or JSONL with original timestamps.
• Works without us. One static binary, no phone-home, no license server. If this repo went dark tomorrow, your instance would not notice.

Don't want to run it? → smolanalytics Cloud

Self-hosting is the free tier, unlimited, forever. The hosted cloud adds an isolated instance per project, your whole team, the morning brief delivered, and scale with zero ops. 14-day full-product trial (no card), then Pro $49/mo (1M events) or Scale $149/mo (10M events), flat $8 per extra million. Overage never locks your dashboard.

The one thing it deliberately does not do

Feature flags, A/B, heatmaps, surveys, a session inspector, cohorts: all ship, all from the same binary. The single deliberate exception is pixel-perfect DOM / video session replay (the screen-recording kind), which needs a heavy recorder and a separate blob store and would break the single-binary model. The event-based session inspector ships instead. Also by design: no multi-node / clustering / HA. Exactly one writer per instance is why it self-hosts in 30 seconds.

Contributing

PRs welcome. Keep it small, correct, and dependency-free (CONTRIBUTING.md). Security: SECURITY.md.

License

MIT, forever. No CLA, no relicense: the business is the hosted cloud, never the license. Use it, fork it, host it, sell hosting of it.