Measuring a team's AI work only works if the team trusts the measurement. So the boundary is architectural, not a policy promise: the collector is built so that raw content cannot reach the cloud at the default level.
Transcripts exist only on the engineer's machine. Every derived stat (tokens, tool health, prompt lengths, session titles) is computed locally by the uploader and shipped as a digest. The cloud never parses raw content.
Access follows roles, enforced at the database layer. Admins read their org. Members read themselves. An engineer always sees their own data; the lead sees the team they run; nobody sees across org boundaries.
Agent Arena is the one exception, and it is yours to make. The arena is a cross-team leaderboard that shows a player only after that player picked a handle and joined. What it shows is the handle, an optional tagline, and counts: agent runtime, sessions, prompts, tool calls, tokens, streaks, how many agents ran at once, and which agent platforms were used. It never shows session titles, project names, prompts or code, and nothing of yours appears until you join. Leaving removes you from the board immediately. A team appears on the team board only when a workspace admin names it, and only opted-in members count toward it.
The collector ships as a public npm package. Read the source and see exactly which events it emits. No compiled blob, no hidden channel.
The installer writes plain hooks into your agent's settings file. Open the file and read what runs on each event. Delete the hooks and collection stops.
Digests are computed on the machine before upload. What travels is the result of the math, not the material it was computed from.
A measurement tool that engineers route around measures nothing. What we collect is deliberately boring: timestamps and counts are enough to draw the lanes, compute leverage, and surface tool health, and they are not interesting to leak.