Features

AI-guided setup

Your first destination goes live in one guided session — without writing a tag, and without giving up control over what gets published.

Guided setup

Example data: Static example state — not live traffic, no real customer data.
  1. Track AI

    example-shop.test runs on Shopify with Cookiebot as consent tool (high confidence). For a shop I suggest view_item, add_to_cart, begin_checkout and purchase. Shall I draft the plan for Meta and Google Ads?

  2. You

    Yes, Meta first.

  3. Track AI

    Pixel id validated. The Conversions API needs an access token — please enter it in the vault card below; it never appears in this chat.

Vault card · Meta access token

Stored encrypted; visible to no one, including the model.

stored

Test event · purchase

Sent through the real pipeline with the vendor's test event code.

accepted by Meta

Publish version 13

Bound to exactly this diff and to you as the approver.

waiting for your approval
  • + destination meta: browser + server
  • + mapping purchase → Purchase (event id, order id)
  • ~ consent: marketing required for meta
Approve and publish

The assistant proposes, tools validate, you approve.

Example session: platform detected, event plan proposed, token stored in the vault, a real test event accepted, publish waiting for your approval.

Every action is a typed tool call you can see

The assistant turns a domain into a working measurement setup: it detects the platform and consent tool, proposes an event plan for your business type, collects public identifiers in chat and secrets in a vault card, sends a real test event and prepares a publish diff you approve.

The assistant never runs free-form code. It calls server-validated tools: detect the platform, draft the event plan, validate a pixel id, send a test event, prepare a diff. Anything irreversible — publishing, rollback, credential rotation — stops at an approval card bound to the exact change you are looking at.

Every action is a typed tool call you can seeTrack AI proposes changes through typed, server-validated tools; your approval, bound to the diff, publishes a signed configuration that the website verifies before loading.Track AIproposesTyped toolsserver-validatedYour approvalbound to the diffSigned configversion 13Websiteverifies signature
Assistant → typed tools → approval card → signed configuration. Secrets go straight into the vault and never pass through the chat.

How it is built

The technical decisions behind the capability — the proof after the benefit.

  1. 01

    Typed tools instead of free-form actions

    Every action the assistant takes is a server-validated tool call with role checks, an audit entry and, for anything irreversible, an approval token bound to the exact diff you saw.

  2. 02

    Secrets never reach the model

    Access tokens go straight into the encrypted vault through a dedicated card or OAuth. The transcript, the model and the browser never see them; a DLP layer redacts pasted secrets and PII.

  3. 03

    Deterministic state machine

    Setup follows an explicit state machine with requirements and evidence for each state. The same states are available as a rule-based wizard when the AI provider is unavailable — nothing depends on a model being online.

Setup by hand versus the guided session

What changes when the assistant drives the setup and you keep the approvals.

Setup by hand versus the guided session
Setup by hand versus the guided sessionManual container setupGuided setup with Track
Event planWritten from memory, one tag at a timeProposed for your business type, edited in chat, validated before publishing
CredentialsPasted into tag fields and visible to everyone with container accessEntered in a vault card or via OAuth; never shown in the transcript
VerificationHope the vendor received somethingA real test event through the real pipeline with the vendor's verdict
PublishingPublish, then discover what changedDiff, recipients and approval first; a signed, versioned bundle after

What you can verify

Product facts you will find in the dashboard, the docs and the audit log.

  • Business type and platform detection with a confidence level
  • Event plan templates for shops, lead generation, SaaS and publishers
  • Public IDs validated against vendor formats
  • Publish only after a diff, recipient list and explicit confirmation

Questions

Can the assistant publish without me?
No. Publishing, rollbacks, credential rotation and destination activation always require your click on an approval card that is bound to the exact change.
Which model is used?
OpenAI Responses API with structured outputs and strict function calling. Model names are configured server-side and verified at start; the UI never hard-codes them.

More capabilities

Built on the same signed configuration and event schema.

Try it on your domain

Create a site, install one snippet and let the assistant configure the first destination in minutes.