Traks

Self-hosted, privacy-friendly web analytics built entirely on Cloudflare.

Traks is a lightweight, cookie-free analytics platform that runs end to end on
Cloudflare's data platform — Workers, Durable Objects, D1, Pipelines, R2 Data
Catalog (Apache Iceberg), and R2 SQL. No servers to manage, no third-party
services in the data path, and no personal data stored.

Open source under the MIT license. Free to run, forever: the only cost is your
own Cloudflare usage, which stays inside the free allowances for most sites.
Install it at traks.dev; read the release notes at
traks.dev/changelog.

Highlights

• Privacy-first — no cookies, no fingerprinting persistence. Visitors are
  counted with a Plausible-style daily-rotating hash
  (HMAC(secret + date, ip + ua + siteKey)); raw IP addresses are never stored.
• Realtime by default — a hot/cold split serves "today" and live views from
  per-site SQLite Durable Objects in milliseconds, with zero ingest delay. A
  WebSocket pushes live visitors, pages, referrers, and city-level map dots to
  the dashboard as they happen.
• Cheap at any scale — history lives in Apache Iceberg on R2 and is queried
  with R2 SQL, edge-cached, and scan-minimized. A side project runs for ~$5/mo,
  5M pageviews/mo for ~$7, and past that about $4.60 per additional million
  events (see cost model).
• Fully self-contained — auth is Better Auth
  on D1 (no auth SaaS), the world map is self-hosted (no tile servers), and the
  dashboard never calls a third party.
• Agent-ready — analytics are exposed to AI agents via MCP/WebMCP tools,
  with bot and agent traffic classified and reported alongside human traffic.
• Tiny tracker — a single t.js script tag, served inline from the edge.

Architecture

Fresh data is served from per-site Durable Objects in milliseconds; history is
served from Iceberg via R2 SQL.

customer site
  └─ t.js tracker (packages/tracker)
       │  POST /api/event
       ▼
collect Worker (apps/platform/collect)            ── site-key auth + timezone (D1)
       │                                    bot filtering, UA/referrer parsing,
       │  dual write                        daily-rotating visitor hash (HMAC)
       ├────────────────────────────┐
       ▼ env.EVENTS.send()          ▼ env.LIVE (SiteLiveStore DO)
Pipelines stream                 HOT PATH: per-site SQLite DO
       │  pass-through pipeline     rolling ~48h event window
       ▼                            zero ingest delay, ms queries
Iceberg sink → R2 Data Catalog      serves: today, realtime
  table `traks.events`                      ▲
  (zstd parquet, 60s roll,                  │
   auto compaction +                        │
   snapshot expiration)                     │
       ▼                                    │
COLD PATH: R2 SQL ◄── api Worker (apps/platform/api) ── Better Auth, D1 metadata
  serves: 7d/30d/90d/1y/all   ▲                today/realtime → DO
  (edge-cached 5-15 min)      │                history → R2 SQL
                              │                (DO failure → R2 SQL fallback)
                 web dashboard (apps/platform/web)

How the pieces fit

• apps/platform/collect — ingest Worker. Validates the site key against
  D1, filters bots, computes the daily-rotating visitor ID, enriches events
  with Cloudflare geo data, then dual-writes: to the Pipelines stream
  (durable system of record) and to the site's SiteLiveStore Durable Object
  (hot path). Each write fails independently.
• SiteLiveStore DO — one SQLite-backed instance per site holding a
  rolling ~48h event window (today plus the previous-day comparison window in
  any timezone). Today/realtime queries run against local SQLite with
  millisecond latency. It is also the realtime push source: the dashboard
  opens one authenticated WebSocket (WebSocket Hibernation API) and receives a
  frame — live visitors, their pages, referrers, countries, and city-level
  coordinates — whenever a pageview changes the picture, plus a 30s tick so
  counts decay as visitors leave. Coordinates exist only in this hot window
  and are never written to Iceberg.
• Pipeline — pass-through INSERT INTO <sink> SELECT * FROM <stream>;
  the stream schema lives in scripts/pipeline-schema.json.
• apps/platform/api — dashboard API. Site/user metadata in D1 (Drizzle);
  historical analytics served by R2 SQL over HTTP, cached at the edge. Also
  serves the dashboard SPA as static assets, so the session cookie is
  first-party by construction.
• apps/platform/web — the dashboard UI, including the live view with a
  self-hosted dotted world map generated from Natural Earth data.
• apps/home — the traks.dev site: landing page, docs, and the install
  wizard. The wizard backend has no database and keeps no record of anyone's
  instance: instances are discovered live from the user's own Cloudflare
  account on each sign-in, and a run's progress lives in a Durable Object
  that wipes itself after a day.
• packages/tracker — the t.js tracking snippet.
• packages/shared — event schema (zod), timezone-aware period math, and
  all R2 SQL query builders.

Bucket keys (date_key, hour_key, week_key) are computed at ingest in the
site's IANA timezone, so dashboard buckets align with the user's local clock.

Repository layout

apps/
  home/            traks.dev site (landing, docs, install wizard)
    api/           home API Worker
    web/           home web app
  platform/        the analytics product
    collect/       ingest Worker + SiteLiveStore Durable Object
    api/           dashboard API Worker (auth, R2 SQL, static assets)
    web/           dashboard SPA
packages/
  tracker/         t.js tracking script
  shared/          event schema, period math, R2 SQL query builders
  eslint-config/   shared lint config
  typescript-config/ shared tsconfig
installer/         release build + upload tooling
scripts/           data-platform provisioning, seeding, tracker inlining

Monorepo managed with Yarn workspaces + Turborepo. Requires Node ≥ 20.

Getting started

Use Traks

You do not need this repository to run Traks. Open
traks.dev/deploy, sign in with Cloudflare, and the
wizard provisions everything into your own account in about two minutes:
both Workers, D1, KV, the R2 bucket with Data Catalog, the Pipelines stream
and Iceberg sink. Updates and removal are one click each at
traks.dev/update and
traks.dev/destroy. traks.dev keeps no record of
your instance; every visit rediscovers it from your account.

Develop Traks

The rest of this section is for working on the platform itself.

Secrets come from Doppler and nowhere else (see Development).
The maintainers' projects are traks-api, traks-collect, and traks-home;
to run the platform locally you need Doppler projects of your own with those
names and the keys listed below.

1. Provision a dev data platform (once per Cloudflare account; reads
CATALOG_TOKEN from Doppler traks-home/prd):

./scripts/setup-data-platform.sh dev

This creates the R2 bucket, enables the Data Catalog with automatic compaction
(128 MB) and snapshot expiration (30 days / keep 5), then creates the stream,
Iceberg sink (60 s roll interval for ~1-minute dashboard freshness), and
pipeline. Paste the printed stream ID into
apps/platform/collect/wrangler.toml.

2. Migrate D1 and start the dev servers:

yarn install
yarn workspace @traks/platform-api db:migrate:dev
yarn dev                                            # collect :5010, api :5011, web :5012, home :5013/:5014

3. Seed test data:

node scripts/seed-events.mjs <SITE_KEY> 500

Useful commands

# Ad-hoc queries (token needs Workers R2 SQL Read)
WRANGLER_R2_SQL_AUTH_TOKEN=<token> npx wrangler r2 sql query \
  "<ACCOUNT_ID>_traks-events-dev" "SELECT COUNT(*) FROM traks.events"

# Catalog / maintenance status
npx wrangler r2 bucket catalog get traks-events-dev

# Pipeline plumbing
npx wrangler pipelines list
npx wrangler pipelines streams list
npx wrangler pipelines sinks list

Warning: never delete objects manually in the catalog-enabled bucket —
data/metadata files under the warehouse prefix are Iceberg table state.

Authentication

Auth is Better Auth running inside the api Worker —
no auth SaaS, no third party. Email + password only; users, sessions, and
credential accounts live in D1.

First-run claim: a fresh instance is unclaimed — /login shows a "create
your owner account" screen, and the first sign-up claims the instance;
sign-ups are rejected server-side after that. The install wizard mints a
one-time CLAIM_TOKEN worker secret and links to /login?claim=<code> so
predictable instance hostnames can't be hijacked.

Recovery (forgot password, no email sending configured): delete the
owner's row in accounts (+ sessions) and re-claim with the same email —
site ownership is re-adopted by email.

What it costs to run

Everything runs inside a Cloudflare Workers Paid plan. Billing for Pipelines,
R2 Data Catalog, and R2 SQL has been live since 3 Aug 2026; each has a monthly
free allowance that most sites never exhaust, so the bill for a small install
is essentially the $5/mo Workers Paid base. Cloudflare bills per event, and
a pageview produces about two (the pageview plus its engagement event), so the
tiers below are stated in both. Rates verified 4 Sep 2026; the same model
drives the calculator on traks.dev.

Above roughly 6M pageviews the bill is dominated by one line: the hot-path
Durable Object writes every event to SQLite, and Cloudflare bills 4 row writes
per event over its lifetime (1 row + 2 index entries on insert, 1 on prune;
measured with cursor.rowsWritten, the figure Cloudflare bills on). That is
$4.00 of the ≈ $4.60 each additional million events costs; Worker requests,
DO requests, Pipelines and R2 together are the remaining cents. For
comparison, hosted analytics vendors publish $16–34 per million pageviews at
their top tiers.

The hot/cold split is what keeps everything else flat: the always-open "today"
dashboard is served by Durable Objects for ~free, historical queries are
minimized to single scans (CASE split for current + previous period
comparisons) and cached for 5–15 minutes, and egress is always $0.

Cloudflare data platform status

R2 Data Catalog, R2 SQL, and Pipelines are still open beta (as of
Sep 2026) but production-trending: pricing is published and billing has been
on since Aug 2026, the catalog has a dedicated dashboard, GraphQL metrics, and
Terraform support, and R2 SQL supports JOINs, CTEs, CASE, window functions,
set operations, exact COUNT(DISTINCT), and ~200 functions.

Known platform gaps this codebase works around: catalog sinks have no
user-defined partition spec and cannot be modified or re-attached to an
existing table (so a sink's roll interval is fixed for the table's lifetime),
stream schemas are immutable, R2 SQL has no timezone conversion and no
metrics dataset for bytes scanned, and every R2 SQL query bills a 10 MB
minimum.

Known limits of this design, from Cloudflare's published numbers:

• One Durable Object per site is the per-site ceiling. Cloudflare rates a
  single object at roughly 200–500 requests/s for operations that write
  storage, and a pageview costs about two DO requests, so one site sustaining
  150–250 pageviews/s (a front-page spike, or roughly 10M+ pageviews/month
  with peaks) saturates its object. The failure is graceful: the write is
  counted in the live_write_failed metric and the dashboard falls back to
  R2 SQL, but "today" undercounts until traffic drops. Sharding a site across
  several objects (keyed by site plus shard, fan-in on read) is the planned
  path if a real site gets there.
• Pipelines beta limits per account: 20 streams, 20 sinks, 20 pipelines,
  and 5 MB/s ingest per stream (about 6k events/s at Traks' record size).
  Each instance uses one of each.
• The Workers Cache API is only functional on custom domains. Wizard
  installs default to workers.dev, where the per-colo cache layer is a no-op
  and the KV result cache does all the work. Instances on a custom domain get
  both layers.

Platform features this codebase relies on:

Abuse guards

Ingest is protected without any per-event database reads: the collect Worker
applies per-site-key burst limits counted per colo, caches site-key auth per
isolate, and accepts events only from the site's registered domain, its
subdomains, and localhost.

Development

yarn dev          # run all apps (collect :5010, api :5011, web :5012)
yarn lint         # lint all workspaces
yarn type-check   # typecheck all workspaces
yarn build        # build all workspaces
yarn check:ci     # everything CI runs: format, lint, build, db, tracker

Secrets come from Doppler only. Nothing reads a secret from the shell
environment or a local file. The dev scripts download each Worker's dev
config (traks-api, traks-collect, traks-home) into a git-ignored
.dev.vars.doppler and hand it to wrangler dev; the release and setup
scripts under installer/ and scripts/ read traks-home/prd. Running any of
them needs doppler login and access to those projects (or Doppler projects of
your own with the same names and keys: VISITOR_HASH_SECRET for collect,
BETTER_AUTH_SECRET and R2_SQL_TOKEN for the api, ADMIN_KEY for home).

Contributing

Issues and pull requests are welcome; see CONTRIBUTING.md
for the setup, the checks to run, and how release notes work. Security issues
go through SECURITY.md, not the public tracker.

License

MIT © 2026 Shivaprasad Manupadi