Add MCP server exposing Glean metadata to AI assistants

[akkomar]

@edugfilho This is a PoC, I'd like to get an early opinion. This implements a simple MCP server for Glean Dictionary. It's not fully compliant with https://github.com/modelcontextprotocol, but supports JSON-RPC over POST which is enough for Claude Code and other tools. Available tools: list_apps, get_app, search_metrics, get_metric, get_ping.
If we decide to productionize this we might want to use the Typescript SDK. Continuing with a simple Js implementation like this is probably fine too.

Why?

While testing SQL generation with AI agents (e.g. in https://github.com/akkomar/mozdata-claude-plugin) we found that the agent having access to Glean metadata can be useful.

Glean Dictionary is a Single Page Application with metrics loaded dynamically via JavaScript. This means that AI web fetchers only see an empty HTML shell, not the actual content. Two workarounds for that were:

1. Directly use https://probeinfo.telemetry.mozilla.org/ (which requires you to know the API in order to write instructions and doesn't contain annotations)
2. Custom MCP server to do above thing: https://github.com/irrationalagent/glean_mcp built by @irrationalagent (with no access to annotations)

Given that it's low effort, I think it would make sense to have an official MCP endpoint in dictionary with support for all the available telemetry metadata.

Local testing

First start the server:

npm install
npx netlify dev

Then add it to Claude:

claude mcp add --transport http glean-dictionary-local http://localhost:8888/mcp

Then in Claude we can search metrics:

❯ What metrics does Firefox Desktop collect about search?

⏺ I'll search for search-related metrics in Firefox Desktop using the Glean dictionary.

⏺ glean-dictionary-local - search_metrics (MCP)(app_name: "firefox_desktop", query: "search", limit: 50)
  ⎿  {
       "metrics": [
         {
     … +534 lines (ctrl+o to expand)

⏺ Let me get more search metrics to provide a complete picture.

⏺ glean-dictionary-local - search_metrics (MCP)(app_name: "firefox_desktop", query: "search", limit: 50, offset: 50)
  ⎿  {
       "metrics": [
         {
     … +530 lines (ctrl+o to expand)

⏺ glean-dictionary-local - search_metrics (MCP)(app_name: "firefox_desktop", query: "serp", limit: 50)
  ⎿  {
       "metrics": [
         {
     … +425 lines (ctrl+o to expand)

⏺ Firefox Desktop collects extensive search telemetry. Here's a summary organized by category:

  1. Search Access Point (SAP) Counts

  Tracks where users initiate searches from:
  - sap.counts - Event recording when a search is initiated from any search access point
  - sap.search_form_counts - Visits to search engine homepages via SAP
  - Entry points tracked: urlbar, searchbar, context menu, about:home, about:newtab, webextensions, system (command line)

  2. SERP (Search Engine Results Page) Telemetry

  - serp.impression - When a SERP is shown to a user
  - serp.engagement - User actions on a SERP
  - serp.abandonment - When user leaves SERP without engaging
  - serp.ad_impression - When ads are detected on a SERP
  - serp.categorization - High-level topic categorization (sports, travel, etc.)

  3. Ad Tracking (by search origin)

  browser.search.withads.* - Counts SERPs displaying ads
  browser.search.adclicks.* - Clicks on ads on SERPs
  browser.search.content.* - In-content search counts

  All broken down by origin: about_home, about_newtab, urlbar, searchbar, contextmenu, webextension, system, etc.

  4. Default Engine Metrics

  - search.engine.default.display_name - Default engine name
  - search.engine.default.engine_id - Engine telemetry ID
  - search.engine.default.provider_id - Engine provider identifier
  - search.engine.default.partner_code - Partner code if applicable
  - search.engine.default.changed - Event when default engine changes
  - Separate metrics exist for private browsing engine (search.engine.private.*)

  5. Newtab Search Metrics

  - newtab.search.issued - Search initiated from newtab
  - newtab.search.enabled - Whether search input is enabled
  - newtab.search.ad.click / newtab.search.ad.impression - Ad interactions from newtab searches
  - newtab.trending_search_* - Trending search widget interactions

  6. Add-on Search Detection

  - addons.search_detection.etld_change_* - Detects when search queries are redirected by extensions

  7. Activity Stream Search Events

  - activity_stream.event_search - User interactions on Activity Stream
  - activity_stream.event_search_handoff - Search handoff events

  8. UI/Navigation Metrics

  - browser.engagement.navigation.searchbar - URI loads from searchbar
  - browser.engagement.navigation.urlbar_searchmode - Loads from urlbar search mode
  - urlbar.persistedsearchterms.* - Search term persistence in urlbar

  The key format for many metrics is <provider>:<tagged|organic> or <provider>:[tagged|tagged-follow-on|organic]:[code|other|none], indicating the search provider and whether it was a tagged (partner) or organic search.

or access ping metadata including [annotations](https://github.com/mozilla/glean-annotations/blob/main/annotations/firefox_desktop/pings/messaging-system/README.md):

❯ show me information about messaging-system ping from firefox desktop

⏺ glean-dictionary-local - get_ping (MCP)(app_name: "firefox_desktop", ping_name: "messaging-system")
  ⎿  ⚠ Large MCP response (~12.7k tokens), this can fill up context quickly
  ⎿  {
       "app_name": "firefox_desktop",
       "name": "messaging-system",
     … +479 lines (ctrl+o to expand)

⏺ Messaging System Ping - Firefox Desktop

  The messaging-system ping captures single events triggered by Firefox's messaging system, including About:Welcome, ASRouter (Activity Stream Router), and other messaging surfaces.

  Key Characteristics
  ┌────────────────────┬──────────────────────────────────────────────────────┐
  │      Property      │                        Value                         │
  ├────────────────────┼──────────────────────────────────────────────────────┤
  │ Includes client_id │ No (uses separate messaging_system.client_id)        │
  ├────────────────────┼──────────────────────────────────────────────────────┤
  │ Send if empty      │ No                                                   │
  ├────────────────────┼──────────────────────────────────────────────────────┤
  │ Bug                │ https://bugzilla.mozilla.org/show_bug.cgi?id=1825863 │
  └────────────────────┴──────────────────────────────────────────────────────┘
  Important Note on client_id

  This ping uses messaging_system_client_id (usually a copy of Legacy Telemetry client ID), not the Glean client ID. When joining with other tables, use metrics.uuid.legacy_telemetry_client_id.

  ---
  Metrics Included (36 total)

  Core Event Data:
  - messaging_system.event - Type of event (IMPRESSION, CLICK_BUTTON, SESSION_END, etc.)
  - messaging_system.action - Action type (IMPRESSION or DISMISS)
  - messaging_system.message_id - Message identifier in Activity Stream Router
  - messaging_system.source - Interaction source (frecent_links, newtab, CFR)
  - messaging_system.ping_type - Event type (cfr, onboarding)

  Context Information:
  - messaging_system.page - Page context (about:newtab, about:home, about:welcome, unknown)
  - messaging_system.event_context - Stringified JSON with additional context
  - messaging_system.event_page, event_reason, event_source
  - messaging_system.event_screen_id, event_screen_index, event_screen_family, event_screen_initials

  Attribution Data:
  - messaging_system.attribution.source, .medium, .campaign, .content
  - messaging_system.attribution.experiment, .variation, .ua
  - messaging_system.attribution.dlsource, .dltoken, .msstoresignedin

  Identifiers:
  - messaging_system.client_id - Telemetry client ID (may be canary ID near data upload pref changes)
  - messaging_system.browser_session_id - Legacy Telemetry browser session ID
  - messaging_system.impression_id - Unique impression identifier
  - messaging_system.bucket_id - Shared name for grouped messages
  - messaging_system.locale - App locale (BCP47)

  CFR-specific:
  - messaging_system.cfr_action - Contextual Feature Recommender action

  ---
  Message Types Captured

  1. CFR (Contextual Feature Recommendations) - Doorhanger recommendations
  2. Moments Pages - Homepage override messages
  3. InfoBar - Messages between toolbars and content
  4. Spotlight - Modal dialog messages
  5. Toast Notifications - OS-level notifications
  6. Feature Callouts - In-page callout messages
  7. Onboarding (about:welcome) - First-run experience screens
  8. Toolbar Badges - Browser toolbar notifications

  Common Event Types
  ┌────────────────────────────┬──────────────────────────┐
  │           Event            │       Description        │
  ├────────────────────────────┼──────────────────────────┤
  │ IMPRESSION                 │ Message was displayed    │
  ├────────────────────────────┼──────────────────────────┤
  │ CLICK / CLICK_BUTTON       │ User clicked             │
  ├────────────────────────────┼──────────────────────────┤
  │ DISMISS                    │ User dismissed           │
  ├────────────────────────────┼──────────────────────────┤
  │ INSTALL / PIN / BLOCK      │ CFR-specific actions     │
  ├────────────────────────────┼──────────────────────────┤
  │ SESSION_END                │ Onboarding session ended │
  ├────────────────────────────┼──────────────────────────┤
  │ TARGETING_EXPRESSION_ERROR │ JEXL targeting error     │
  └────────────────────────────┴──────────────────────────┘
  Privacy Handling

  - Release channel: Excludes client_id and uses bucket_id instead of message_id
  - Private browsing: Same as release channel
  - Prerelease/experiments: Includes full client_id and message_id

[edugfilho]

I've never implemented an MCP server before, but this is an amazing idea.