Usage & Cost
The Usage & Cost page gives coding-agent channels a single view of what your team is spending on AI coding agents and how widely they're adopted — across Claude Code, Codex, and Cursor, attributed per workstation.
The tab appears in the channel sidebar for coding-agent channels.
What it shows
Spend and Adoption are scoped to a time range you choose (7, 30, or 90 days) with the segmented control at the top. Governance below them is live — the current fleet and policy posture, not a windowed metric.
Spend
- Modeled cost — token usage priced with a built-in per-model rate table. Cost is a model over token counts; no prompt or code content is involved.
- Tokens — total volume, broken out into input, output, and cache (cache-read plus cache-creation) tokens, each priced at its own rate.
- By model — tokens and cost per model, busiest first. A model that isn't in
the pricing table shows its tokens with a cost of "—" rather than a guess, and
the headline cost is marked as a lower bound (
≥) when that happens. - Tokens by day — a trend bar per day in the range.
Adoption
- Active workstations — the distinct machines that reported activity.
- Sessions — the distinct coding-agent sessions in the range.
- Runtimes — which agents (Claude Code, Codex, Cursor) are in use, and how many workstations and sessions each covers.
- Tool calls — the mix of tools the agents invoked (Bash, Edit, Read, …).
- Sessions by day — an activity trend per day.
Governance
Governance joins your fleet's enrollment with the policy the channel serves — a read over state the fleet already holds, not new data collection.
- Enrolled — coding-agent workstations known to the channel, and how many are active right now.
- Config drift — workstations whose applied configuration has drifted from the deployed policy (they're not running what you deployed).
- Enforcement mode — the split between cooperative workstations (running the user-mode supervisor) and enforced ones (the managed root service).
- Served policy — the deployed redaction tier and what it lets leave the workstation, plus whether usage metadata forwards to Ampbase and whether redacted events forward to your own SIEM. It also shows whether edge identity-drop is on: when enabled in the policy, the supervisor omits developer, team, region, host, and the workstation's instance identifier from usage records at the workstation, so no per-developer or per-workstation attribution is ever retained — data minimization at the edge, for teams that need it. Spend and adoption still work; per-workstation breakdowns do not.
Adoption without spend
It's normal to see tool activity with zero tokens shortly after enrolling. In cooperative mode the supervisor applies its hooks to sessions that are already running, but token telemetry only wires in for sessions started after enrollment. Until existing sessions cycle, you'll see adoption climb before spend does — that's expected, not a gap.
Privacy
Everything on this page is computed from redacted usage metadata — token counts, tool names, session and model identifiers. Prompts, code, diffs, and command output never leave your supervisor-managed devices (workstations and CI runners alike) to Ampbase; if you route that content anywhere, it goes to your own SIEM. The one exception is customer-directed: cloud coding agents you point straight at your ingest endpoint arrive without device-side redaction — see cloud coding-agent OTLP intake. Nothing on this page is computed from that content either way.