What's new.

All notable changes to this project are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

While the major version is 0, breaking changes may ship in a minor release.

Added

• In-app AI chat assistant — an optional chat widget that answers questions about your analytics using the same read-only tools as the MCP server. Bring your own API key for one of six providers (Anthropic, OpenAI, Gemini, Groq, OpenRouter, or NVIDIA NIM); if multiple keys are set, the first available wins in that priority order. Provider, model, and base URL are all configurable via environment variables — see .env.local.example. Disabled by default (no keys set).
• Detailed error tracking — error detail pages now show affected-user and affected-session counts, a daily occurrence timeline, and breakdowns by browser/OS/device/country, backed by a new per-occurrence error_occurrences table (capped at 500 stored occurrences per error per day; prune with the new ERROR_OCCURRENCE_RETENTION_DAYS). Full, multi-frame stack traces are now resolved through source maps (previously only the top frame). Breadcrumbs — console logs, clicks, network requests, and navigations in the moments before an error — are captured automatically and shown on the detail page. Errors now carry a filterable, structured type (e.g. TypeError), and a new window.__DIY_CAPTURE_EXCEPTION__(error, { level, tags, extra }) lets you report caught exceptions (e.g. from your own React error boundary) with custom severity and context. New MCP tool: get_error_occurrences.

Changed

• Errors: error groups whose message is exactly Script error. with no stack or source location — the browser's redaction of uncaught errors thrown by a cross-origin <script> without crossorigin="anonymous" and a matching CORS header — are now tagged cross-origin in the list, with an explanation and fix in the detail view instead of a dead-end "No stack trace captured."

Fixed

• The chat pill's response bubble and "thinking" indicator now hide whenever the pill loses focus, instead of staying visible after you click away.
• The chat pill is now usable on mobile: the input no longer triggers iOS Safari's auto-zoom on focus, the collapsed pill fits narrower screens, and it no longer sits flush against the home-indicator bar on notched devices.

Changed

• Analytics dashboard read performance — the Overview dashboard is substantially faster, especially on projects with large event/pageview volumes. Verified against a seeded 200K-pageview dataset: full-payload response time dropped from ~16.5s to ~1.95s, and the slowest single panel query from ~21.7s to ~1.6s. On typical project sizes, time to first meaningful paint dropped from ~4.6s to under 400ms.
  • Fixed a chart-bucketing routine that was doing redundant timezone conversions per pageview per bucket instead of once per request.
  • Batched several breakdown queries (country/browser/os/device/source/ campaign, and web vitals per-dimension) into single grouped queries instead of scanning the same window separately for each.
  • Moved session/bounce-rate/duration and top-pages engagement metrics from in-app JavaScript aggregation to SQL, so the response no longer grows with pageview volume.
  • The dashboard now fetches in three parallel segments (core metrics, breakdowns, insights) instead of one blocking request, so metric tiles and the main chart can render before slower panels finish loading.

Fixed

• The Profile "System" storage card could report a wildly inflated "days until storage is full" estimate after a large delete (e.g. data retention cleanup), because it derived average row size from total table size including reclaimable space rather than from actual row content. It now uses Postgres's column-width statistics, which aren't affected by table bloat.

Added

• MCP server — every deployment now serves an MCP endpoint at /api/mcp, so an AI assistant (Claude Desktop, Cursor, or any other MCP-compatible client) can query analytics, funnels, retention, errors, and more directly. Ships with 14 read-only tools; nothing can be created, edited, or deleted through MCP yet.
• API keys — a new Profile → API Keys page for creating/revoking personal bearer tokens used to authenticate MCP clients. Keys are user-scoped (not project-scoped); every project-scoped tool call re-checks access on every invocation, so a revoked workspace membership takes effect immediately.
• Multiple domains per project — projects can now authorize one or more additional domains (e.g. a *.vercel.app preview kept alongside a custom domain) from Settings → Tracking, on top of the existing primary URL. Authorizing a domain also authorizes its subdomains, one-directionally.

Changed

• The Profile page is now a sidebar layout (Account / Workspaces / API Keys / Storage) instead of a single scrolling column of cards.
• Renamed the Page Flow tab to Journeys throughout the app and docs.
• docs/features.md and docs/integration.md were substantially rewritten: reorganized to match the app's actual sidebar order, and filled in previously-undocumented features (Explore, Journeys, Error tracking, Segments, multi-domain authorization, MCP server).

Added

• Explore — an ad-hoc session query builder: combine attribute (country, browser, device, source, page, UTM, OS, city), "visited page", and "fired event" conditions with AND/OR to find matching sessions without a predefined funnel or goal.
• Errors: a "new" badge for errors first seen within the last 24h, and a release filter/breakdown (grouped counts per release tag) on the Errors page.
• Resource-load failures (blocked/failed script, image, and link loads) are now captured alongside JS exceptions and unhandled rejections, via a capture-phase listener — previously only thrown errors were tracked.

Fixed

• Explore queries failed with a driver-level type error because date-range bounds were interpolated as raw Date objects into hand-written SQL instead of going through drizzle's typed comparison operators.

Added

• Error tracking — uncaught exceptions and unhandled promise rejections are now captured client-side (sampled, capped per page load), grouped by a message+stack fingerprint, and shown on a new Errors tab with severity, regression detection (a resolved error reoccurring reopens it), and a resolve/reopen workflow. Stack frames can be resolved to original source via an on-demand source map fetch (no server-side map storage required).
• Page Flow — a page-to-page transition diagram (top sources → top destinations) computed from session-ordered pageviews.
• Segments — visitors grouped into four recency/frequency segments (Active & frequent, Active & occasional, Lapsing, Dormant). Anonymous by default; call the new window.identify(uid) tracker API to segment real people instead of browsers.
• Anomaly alerts — a new "Anomalous vs. trailing average" alert threshold compares the live value against a 14-day rolling mean/stddev from daily rollups, instead of a fixed number.
• Goal- and funnel-based alerts — alerts can now target a goal's or a funnel's conversion rate, not just a site-wide metric.
• Slack, Discord, and PagerDuty alert channels, alongside the existing generic webhook.
• Web Vitals: p50 shown alongside p75, plus a breakdown by page, country, device, and browser (previously site-wide only).
• Resource timing — the slowest page assets (script/image/fetch/etc.) are now sampled and surfaced as a waterfall-style panel.

Changed

• Breaking: migrated the persistence layer from MongoDB/Mongoose to PostgreSQL via Drizzle ORM. MONGODB_URI/MONGODB_DATABASE are replaced by DATABASE_URL. Existing self-hosted deployments must provision a Postgres database, run npm run db:migrate, and optionally npm run db:migrate-from-mongo to import existing data — see docs/migrating-from-mongodb.md.
• MONGODB_STORAGE_CAP_MB is renamed DATABASE_STORAGE_CAP_MB.
• App version is now shown in the sidebar footer, sourced from NEXT_PUBLIC_APP_VERSION.

Added

• Core analytics dashboard: traffic, top pages, sources, campaigns, countries, devices, browsers, and a live visitor count.
• Real-time and historical views with per-project timezone support.
• Goals and multi-step funnels.
• Retention cohorts.
• Threshold-based alerts via webhook.
• Public, read-only dashboards with a shareable link.
• Workspaces with role-based member access (viewer / member / admin / owner).
• Self-hosted tracking script (/api/tracker.js) with automatic SPA pageview tracking, custom events, Core Web Vitals, and a visitor opt-out API.