other or custom. GitHub announced the change on October 22, extending September’s customization reporting.For administrators, the useful distinction is between knowing that developers use something custom and knowing which internal tool deserves attention. But clearer labels do not turn these reports into an exact adoption census.
What changes in the reports
The update affects five arrays:
totals_by_skilltotals_by_custom_agenttotals_by_mcptotals_by_slash_cmdtotals_by_plugin
Each list expands from five to up to 25 entries per record, with remaining activity combined under a trailing other entry. The calculations behind interaction_count and distinct_*_use_count remain unchanged.
GitHub’s documentation places customization metrics in enterprise and organization reports: per-user 1-day and 28-day reports, aggregated 1-day reports, and individual day_totals records within aggregated 28-day reports. These customization breakdowns are API-only, not available in the usage dashboard.
That makes this principally a reporting-pipeline improvement: teams consuming the API gain more useful identifiers without a new counting methodology.
What an “interaction” actually means
GitHub’s metrics reference makes an important distinction: these are event counts, not user counts, and each category measures a different event.
| Category | What interaction_count measures |
|---|---|
| Skills | Skill invocations |
| Custom agents | Agent starts |
| MCP servers | Successful or failed connection and reconnection attempts |
| Slash commands | Command invocations |
| Plugins | Skill invocations associated with a plugin |
MCP tool calls through an already connected server do not increase its count. Plugin interactions also appear in skill totals, so adding the two categories together double-counts that activity.
Consequently, a busy MCP entry is not necessarily evidence of a productive integration. GitHub’s interpretation guidance recommends checking configuration and rollout conditions because repeated connections can reflect failures or reconnections—not sustained tool use.
The practical lesson: compare like with like. An agent start and an MCP connection attempt are not interchangeable units of developer value.
Client versions determine name visibility
GitHub lists these reporting boundaries:
- Copilot CLI: interactive sessions and prompt mode are covered.
- Prompt mode, Copilot app and ACP integrations, including JetBrains: require CLI 1.0.97 or later.
- Copilot SDK and SDK-powered IDE agents: require SDK v1.0.21 or later.
- Built-in IDE chat telemetry: does not report names.
Before treating a missing name as an adoption problem, establish which client and reporting path generated the activity.
Distinct counts are not headcounts
The distinct_*_use_count fields measure different identifiers with recorded activity. At organization or enterprise scope, an identifier counts once regardless of how many developers used it. Aggregate distinct counts therefore cannot be reconstructed by adding per-user distinct counts. They also include identifiers outside the displayed arrays.
For example, if ten developers use the same custom agent, that represents one distinct agent at organization scope—not ten. This is a measure of customization variety, not audience size.
GitHub also warns against treating an absent array entry as zero use or using these limited breakdowns to calculate exact adoption. Its interpretation page still describes the earlier five-entry limit; the October announcement explicitly increases that limit to 25. The visibility improves, but the completeness caveat remains.
An administrator’s review checklist
Access requires an eligible owner, enterprise billing manager, or custom role with the relevant Copilot metrics permission, plus an enabled Copilot usage metrics policy.
When evaluating a rollout:
- Check reporting coverage first. Outdated clients or blocked telemetry can undercount activity.
- Review names before distributing exports. Labels are trimmed to 128 characters; path-like names remain grouped.
- Separate historical visibility from adoption. Names appear for report days processed after release; earlier records retain generic labels.
- Pair observed activity with developer feedback. GitHub recommends comparing multiple periods and investigating unexpected MCP connection patterns before changing enablement plans.
The payoff is better evidence for deciding which internal customizations to document, support or investigate. The API can now put names to more activity—but administrators still need to ask what that activity actually represents.
References
- Custom skills, agents, and MCP servers in usage metrics API GitHub Changelog · 2026-10-22T22:19:17+00:00
- Data available in Copilot usage metrics - GitHub Docs docs.github.com
- Interpreting usage and adoption metrics for GitHub Copilot - GitHub Docs docs.github.com