Skip to content
Türkçe

skill + mcp

Analytics your agent ships with ads denied

A container loaded from Google's host, a team plan stored on the user or an email sent as the user id is found only after release. The magic_analytics skill teaches your agent the provider order, the identity rules and the consent wiring before it writes a line, and the MCP server lets it look a signature up in the live docs.

npx skills add anilcancakir/ai --skill magic-analytics

run in your project other editor

works with

  • Claude Code
  • Cursor
  • OpenCode
  • VS Code Copilot
  • Codex CLI

Ask for the feature, not the API.

Three things people ask for when a Magic app gets analytics. The skill turns each one into the calls and the config magic_analytics expects.

  • send an event when a monitor is created, tagged with the team plan

    Calls Analytics.identify with the user's UUID, puts the plan in setDefaultParams so it rides on every event, then sends monitor_created with Analytics.track.

    Reaches for

    • Analytics.identify
    • Analytics.setDefaultParams
    • Analytics.track
  • load GA4 on web from our own domain, not from Google

    Sets GTM_CONTAINER_ID and GTM_GATEWAY_PATH, proxies that path to GTM-XXXXXXX.fps.goog and keeps send_page_view off in the container, because the driver pushes each page view itself.

    Reaches for

    • GtmWebDriver
    • analyticsConfig
    • Env.filled
  • let users opt out of analytics and remember the choice

    Persists the choice in the app and calls Analytics.setOptOut during boot, before the first page view, because the package holds the opt-out in memory only.

    Reaches for

    • Analytics.setOptOut
    • Analytics.isOptedOut
    • Analytics.reset

Where agents slip, and what the skill writes instead.

Each of these passes a quick run and shows up later as wrong GA4 data or a lost user. The skill writes the second version.

A team plan stored as a user property

setUserProperties is user-level, so a team switch leaves the old team on the user. The skill puts team facts in setDefaultParams, which replaces the set and rides on every event.

without the skill
// The old team stays on the user after a switch.
await Analytics.setUserProperties({'team_plan': teamPlan});
with the skill
// Replaced per team and sent with every event.
await Analytics.setDefaultParams({'team_plan': teamPlan});

An email as the GA4 user id

identify refuses an id the scrubber would change, so an email never becomes the user id and the user stays anonymous. The skill passes the user's UUID.

without the skill
Future<void> onSignedIn(String email) async {
  // Refused: the scrubber masks an email.
  await Analytics.identify(email);
}
with the skill
Future<void> onSignedIn(String userUuid) async {
  await Analytics.identify(userUuid);
}

Adding the page view observer by hand

MagicRouter refuses an observer once it has built its routerConfig. The skill registers AnalyticsServiceProvider in providers, after error reporting, and its boot() adds the observer in time.

without the skill
// Too late once the router has built its config.
MagicRouter.instance.addObserver(AnalyticsNavigatorObserver());
with the skill
final List<Object> providers = [
  // ...your other providers, error reporting first
  (app) => AnalyticsServiceProvider(app),
];

Three pieces, one outcome.

The skill shapes what your agent writes, the MCP server checks it and llms.txt fills the gaps.

  • Skill

    edit time

    Teaches your agent the provider order, the identity and default parameter rules and the consent wiring before it writes the first line.

    See the install paths
  • MCP server

    runtime

    Lets your agent look an API up in the live docs while it edits, so it never guesses a signature.

    Connect a client
  • llms.txt

    on demand

    A plain-text map of every docs page, for agents that read it once and keep it in context.

    Open llms.txt

Three paths. Pick one for your editor.

Each path covers this skill; the plugin and the installer cover the whole registry.

  • Claude Code plugin

    Adds the registry as a marketplace, then installs its plugin.

    # add the marketplace

    /plugin marketplace add anilcancakir/ai

    # install the plugin

    /plugin install anilcancakir@anilcancakir-marketplace

    Installs every skill in the registry, not only this one.

  • npx skills recommended

    The universal CLI. It works with the editors the skills package supports.

    # this project

    npx skills add anilcancakir/ai --skill magic-analytics

    # every project

    npx skills add anilcancakir/ai --skill magic-analytics -g

    The same command the landing page shows.

  • Universal installer

    Detects every supported editor and sets each one up the way it supports: skills, rules, commands or printed steps. It also adds the docs MCP server config.

    # clone and run

    git clone https://github.com/anilcancakir/ai.git && cd ai && bash scripts/install.sh --global --with-mcp

    Covers the whole registry, not only this skill.

Connect the MCP server.

One endpoint at mcp.anilcancakir.com/magic_analytics, six clients. Pick yours and paste the config.

What your agent asks the server

  • search_docs "AnalyticsServiceProvider boot observer"

    What the provider wires on boot and why MagicRouter refuses an observer once its config is built.

  • search_docs "GtmWebDriver gateway_path"

    How the container loads lazily from the first-party path and what each call pushes to the data layer.

  • search_docs "UrlScrubber scrub"

    Which query keys, route params and email shapes are masked, and which inputs the facade scrubs.

Claude Code

terminal
claude mcp add --transport http magic_analytics https://mcp.anilcancakir.com/magic_analytics

Cursor

.cursor/mcp.json
{
    "mcpServers": {
        "magic_analytics": {
            "type": "http",
            "url": "https://mcp.anilcancakir.com/magic_analytics"
        }
    }
}

OpenCode

opencode.json
{
    "mcp": {
        "magic_analytics": {
            "type": "remote",
            "url": "https://mcp.anilcancakir.com/magic_analytics"
        }
    }
}

VS Code Copilot

.vscode/mcp.json
{
    "servers": {
        "magic_analytics": {
            "type": "http",
            "url": "https://mcp.anilcancakir.com/magic_analytics"
        }
    }
}

Codex CLI

~/.codex/config.toml
[mcp_servers.magic_analytics]
url = "https://mcp.anilcancakir.com/magic_analytics"

Antigravity CLI

~/.gemini/config/mcp_config.json
{
    "mcpServers": {
        "magic_analytics": {
            "serverUrl": "https://mcp.anilcancakir.com/magic_analytics"
        }
    }
}

Frequently asked

What does plugin:install set up, and what is still mine to do?

It publishes lib/config/analytics.dart, registers AnalyticsServiceProvider in lib/config/app.dart and adds analyticsConfig to the config factories. It writes no native file and cannot set up a web server, so the native keys and the gateway proxy are yours.

Why does the web driver need a gateway path?

So the container is never requested from googletagmanager.com. The web driver needs both GTM_CONTAINER_ID and GTM_GATEWAY_PATH, a path on your origin proxied to GTM-XXXXXXX.fps.goog; without them it is the null driver.

Does the skill also teach GA4, GTM or Firebase?

No. It covers magic_analytics only. The Magic framework is the magic-framework skill, and GA4, GTM and Firebase are Google's own documentation; the MCP server looks magic_analytics APIs up while the agent edits.

Why do opted-out web users still show up in GA4?

The container has to honour the consent update: require analytics_storage on both the Google tag and the GA4 Event tag, or a denied tag still sends cookieless pings. The opt-out itself lives in memory, so set it again on every start.

Which AI coding assistants are supported?

The skill installs into Claude Code, Cursor, OpenCode, VS Code Copilot and Codex CLI. Any client that speaks MCP over HTTP or stdio can use the server.

Where do I report a mistake?

Open an issue on the Magic Analytics repository on GitHub with the prompt, what your agent wrote and the docs page that says otherwise.