I built AI-native analytics that lives inside your AI coding assistant — here's what 3 months of dogfooding taught me
Hey IH! I'm the solo founder of Amami (https://amami.dev) – AI-native web analytics that lives inside your AI coding assistant (Cursor, Claude Code, Codex).
Instead of opening a dashboard, you ask your editor "how was traffic this week?" and it answers with real numbers from your site. The analytics layer is built for agents, not for dashboards.
Why I built this: I ship side projects with AI coding tools constantly. After deploying, I wanted to know if anyone was visiting – but every analytics tool was a dashboard I never opened. The data existed; I just never looked at it. The gap wasn't analytics, it was access: the questions had answers, but asking them required a detour to a website I had no reason to visit.
What's real right now (not a roadmap):
- One-command install: npx -y amami-analytics-mcp@latest setup --write — browser auth, site registration, tracking snippet, client config, all in one flow
- 6 read-only MCP tools your AI calls: list sites, read stats, inspect trends, top pages, active visitors, trace sources
- Natural language Q&A with evidence: the AI shows you which pages and date ranges it looked at
- Read-only by default; write tools require explicit opt-in; credentials live locally, never in chat
- Free tier: 100K events/month, 5 websites, 50 MCP calls/day
What dogfooding taught me (the honest part):
- Asking questions beats browsing charts – not because answers are better, but because you actually act on them. Dashboards are passive consumption; questions force a decision.
- Launch-day traffic ≠ launch-week value. PH sent a burst of curious visitors; a small niche community sent fewer but converted better. Channel quality ≠ channel volume.
- "Why did traffic spike?" is the question that keeps giving. Every spike is either a bug, a bot, or a marketing channel doing something right – and the AI can tell you which.
What I'm working on next: turning the growth layer into structured workflows – daily growth summaries, launch monitoring ("what changed after I shipped?"), and GTM playbooks that turn a data question into an experiment with a measurable outcome. Analytics that helps you sell, not just measure.
We also just opened GitHub Discussions (https://github.com/april-jk/amami/discussions) for anyone building with analytics + AI – troubleshooting, feature ideas, case studies. Come say hi.
Curious from other IH folks: do you actually open your analytics dashboard regularly, or is it the most-bookmarked-never-visited tab in your browser? I suspect the latter is more common than we admit.
The dogfooding result points to an access problem, not a reporting problem. I’d measure the behavior change directly: for the same user, compare how often they ask a question and take a follow-up action versus how often they open a dashboard and do anything within the next day. That would separate “this is convenient” from “this changes decisions.” Your launch-day versus launch-week observation is a useful warning too—I'd segment channels by second-week return rate and the first meaningful action, not just visits. One question I’d test next: when the assistant surfaces a spike, do users change distribution, onboarding, or the product itself? The answer may tell you whether Amami is becoming analytics access or a real growth workflow.
The interesting signal here is that the dashboard may not actually be the product problem — the access pattern might be. If asking a question inside the coding workflow consistently leads to more action than opening analytics separately, that feels like a much bigger shift than simply adding an MCP interface. Curious how consistently you've seen that behavior outside your own dogfooding.
The 'dashboards are passive consumption' line is the one that matters. I shipped a small tool for myself once, same pattern — data existed, I never looked at it, because looking at a chart wasn't an action. The moment something asks me a question ('why did traffic spike?') it forces a decision, and that's when I actually change something. Also seconding the launch-day vs launch-week point: one burst of curious visitors told me almost nothing, a handful of users who came back next week told me everything. Channel quality beats channel volume every time.