Open source - MIT licensed
GTM implementation,
done by AI in 6 steps
A local web app that runs Claude on your codebase to audit tracking gaps, standardise your DOM, plan your GA4 events, write the code, test it, and generate the report. End to end.
6 steps. One pipeline. Zero guesswork.
Each step runs Claude on your actual codebase. You review, approve, and move forward. The pipeline does not proceed without your say-so.
01Audit
Know exactly what is and is not tracked
Claude scans every file in your codebase and maps all trackable elements - buttons, forms, links, interactions - against what is currently instrumented. You get a prioritised list of gaps before writing a single line of code.
Coverage percentage by page
Untracked elements by category
P0 / P1 / P2 priority recommendations
02Prepare Elements
Give every trackable element a stable identity
Adds data-gtm-id attributes and js-track classes to your HTML elements so GTM triggers have a reliable, consistent hook. Idempotent - safe to re-run without duplicating attributes.
DOM-standardised source files
Per-page diff view before applying
Conflict detection for existing attributes
03Strategy
A tracking plan built for your business goal
You answer four questions about your business goal, conversion action, and audience. Claude produces a complete GA4 event taxonomy with parameter schemas, mapped to the audit findings. Requires your approval before implementation begins.
Full GA4 event taxonomy
Parameter schemas per event
Approval gate - nothing runs without sign-off
04Implementation
Code changes and GTM container updates in one pass
Adds dataLayer.push() calls to your source files and creates all GTM variables, triggers, and tags, then publishes the container version. Platform-aware: uses googleapis on Windows, gtm-cli on Mac/Linux. Re-run safe: never duplicates work already done.
dataLayer events in source files
GTM variables, triggers, tags created
Container version published
05Testing
Three-tier validation before you call it done
Static analysis across three tiers: dataLayer.push() calls in source files, GTM container configuration, and GA4 naming conventions. Validates every event from the tracking plan and flags anything that does not match.
Pass / fail per event
GTM trigger and tag validation
GA4 naming rule checks
06Reporting
Implementation report + recommended GA4 reports
Generates a concise implementation report covering coverage improvement, events implemented, files modified, and test results. Also surfaces recommended GA4 Exploration reports tailored to your tracking plan.
Markdown implementation report
Recommended GA4 funnel reports
Custom dashboard suggestions
Built for how tracking actually works
Not a prompt wrapper. A structured pipeline with real guardrails.
Approval gate built in
Strategy output requires explicit approval before implementation runs. Nothing gets deployed without sign-off.
Safe to re-run
Every step is idempotent. Re-running never duplicates attributes, events, or GTM resources. Pick up exactly where you left off.
Runs on your machine
No cloud, no SaaS fees, no data leaving your environment. Runs at localhost using your own Claude subscription.
Parallel strategy variants
Strategy step can run multiple Claude workers in parallel to produce and compare tracking plan variants before you choose one.
Multi-project support
Switch between projects without losing progress. Each project keeps its own session, outputs, and approval history.
Fully open source
MIT licensed. Self-host, fork, extend. All 6 step prompts are plain markdown files you can read and edit.
What it runs on
Requires Node.js 20+ and the Claude Code CLI installed and logged in. No Anthropic API key needed - runs on your Claude subscription.
Get started in minutes
Clone, install, run. The database is created automatically on first launch.
#Clone the repo
$git clone https://github.com/aimonk2025/tag-manager-engine
$cd tag-manager-engine
#Install dependencies
$npm install && cd app && npm install && cd ../server && npm install && cd ..
#Start the app
$npm run dev
Then open localhost:5173, create a project, point it at your codebase, and run step one.