Product · Real-time analyticsDraft
See where the tokens go, minutes after they go there.
A lightweight hook installs in one paste and reports every turn: tokens, model, tool use, and which repo, branch and ticket the work was for. Prompts and code never leave the machine, only the numbers do. Joined with your tickets and pull requests, that becomes cost per PR, code quality, time saved and the full cycle time from ticket to merge, per developer and per workflow.
What the dashboard shows Sample data
human-paced machine-paced · every number is per repo, team or developer
Outcomes
Cost per PR
Model spend per merged pull request
$8.70-41% vs 12 wk ago
$8.70 week of Nov 2
Table
| Week | USD |
|---|---|
| Aug 17 | $14.80 |
| Aug 24 | $13.90 |
| Aug 31 | $15.20 |
| Sep 7 | $12.60 |
| Sep 14 | $11.90 |
| Sep 21 | $12.40 |
| Sep 28 | $10.80 |
| Oct 5 | $10.10 |
| Oct 12 | $10.60 |
| Oct 19 | $9.40 |
| Oct 26 | $9.10 |
| Nov 2 | $8.70 |
Code quality
PRs merged with no rework round after human review
81%+33% vs 12 wk ago
81% week of Nov 2
Table
| Week | % |
|---|---|
| Aug 17 | 61% |
| Aug 24 | 63% |
| Aug 31 | 60% |
| Sep 7 | 66% |
| Sep 14 | 69% |
| Sep 21 | 71% |
| Sep 28 | 70% |
| Oct 5 | 74% |
| Oct 12 | 76% |
| Oct 19 | 75% |
| Oct 26 | 79% |
| Nov 2 | 81% |
Time saved
Human hours saved per week against the pre-agent baseline
49 h+172% vs 12 wk ago
49 h week of Nov 2
Table
| Week | hours |
|---|---|
| Aug 17 | 18 h |
| Aug 24 | 22 h |
| Aug 31 | 21 h |
| Sep 7 | 27 h |
| Sep 14 | 31 h |
| Sep 21 | 30 h |
| Sep 28 | 36 h |
| Oct 5 | 38 h |
| Oct 12 | 41 h |
| Oct 19 | 40 h |
| Oct 26 | 46 h |
| Nov 2 | 49 h |
By agent · last 12 weeks
Cost per agent
Model spend per merged pull request, by agent
$2.86orchestrator · lowest of 10
Reviewer cost per issue found
Model spend per confirmed finding, by reviewer agent
$1.90code-reviewer · lowest of 8
Reviewer performance
Confirmed issues found in the last 12 weeks, by reviewer agent
148code-reviewer · highest of 8
Cycle time, ticket to merge
Ticket pickup
Average time from ticket ready to work started
9.5 h-69% vs 12 wk ago
9.5 h week of Nov 2
Table
| Week | hours |
|---|---|
| Aug 17 | 31 h |
| Aug 24 | 28 h |
| Aug 31 | 29 h |
| Sep 7 | 24 h |
| Sep 14 | 22 h |
| Sep 21 | 19 h |
| Sep 28 | 18 h |
| Oct 5 | 15 h |
| Oct 12 | 14 h |
| Oct 19 | 12 h |
| Oct 26 | 11 h |
| Nov 2 | 9.5 h |
Pickup to PR
Average time from work started to pull request opened
4.1 h-55% vs 12 wk ago
4.1 h week of Nov 2
Table
| Week | hours |
|---|---|
| Aug 17 | 9.2 h |
| Aug 24 | 8.6 h |
| Aug 31 | 8.9 h |
| Sep 7 | 7.4 h |
| Sep 14 | 6.8 h |
| Sep 21 | 6.1 h |
| Sep 28 | 5.9 h |
| Oct 5 | 5.2 h |
| Oct 12 | 4.9 h |
| Oct 19 | 4.6 h |
| Oct 26 | 4.3 h |
| Nov 2 | 4.1 h |
PR to first review
Average wait from pull request opened to first human review
74 min-65% vs 12 wk ago
74 min week of Nov 2
Table
| Week | minutes |
|---|---|
| Aug 17 | 210 min |
| Aug 24 | 190 min |
| Aug 31 | 205 min |
| Sep 7 | 160 min |
| Sep 14 | 150 min |
| Sep 21 | 135 min |
| Sep 28 | 120 min |
| Oct 5 | 110 min |
| Oct 12 | 95 min |
| Oct 19 | 90 min |
| Oct 26 | 82 min |
| Nov 2 | 74 min |
First review to merge
Average time from first human review to merge
4.6 h-60% vs 12 wk ago
4.6 h week of Nov 2
Table
| Week | hours |
|---|---|
| Aug 17 | 12 h |
| Aug 24 | 11 h |
| Aug 31 | 11 h |
| Sep 7 | 9.4 h |
| Sep 14 | 8.6 h |
| Sep 21 | 8.1 h |
| Sep 28 | 7.2 h |
| Oct 5 | 6.6 h |
| Oct 12 | 6.1 h |
| Oct 19 | 5.4 h |
| Oct 26 | 5 h |
| Nov 2 | 4.6 h |
From a turn to a chart
Machine · runs without you
Install
One paste per developer. An admin creates a project key; hooks register with the agent.
- Session start, tool calls, permission requests, sub-agent stops, session end
- Works with Claude Code today, other agents in a simpler view
- One switch turns it all off
machine runs human reviews
What you get
Cost and tokens
Spend by model, skill, workflow step and team, with cache hit rate and a monthly projection.
Efficiency
Tokens per turn, lines per dollar, context size, compactions and tool outcomes.
Cycle time, stage by stage
Ticket pickup, pickup to PR, PR to first review, and first review to merge, each as its own trend.
People and teams
Spend per developer, who is getting the most from the tools, and who has not started.
Privacy by construction
Hooks send counts and identifiers, never prompts, responses or code.
Every major coding agent
Claude Code gets the full dashboard today; Codex, Cursor, Gemini and Grok report into a simpler view.
Questions
How real-time is real-time?
Events arrive per turn while the developer works. Dashboards show them within minutes, not seconds; this is not a live stream and does not need to be.
Does this replace the vendor billing page?
It reconciles against it. Vendor billing tells you the total; this tells you which repo, ticket, workflow step and reviewer the total came from.
Where does the two-clock number come from?
Human time is the gaps between agent turns where a person was reading or deciding; machine time is everything else. Both are derived from the same per-turn events.