For teamsWhat your agents cost and shipped.
Every coding agent on every machine: what it is doing now, what it cost per pull request, however you pay.
Spend, week 39
$2,310
API list price
Merged PRs
58
read from GitHub
Per merged PR
≈$33
estimate · $420 unlinked
Running now
6
on 5 machines · 1 loop
By repositoryper merged PR · PRs shipped
- payments-core$35 / PR19 PRs
- web-client$29 / PR16 PRs
- data-pipeline$41 / PR9 PRs
- admin-panel$26 / PR9 PRs
- infra$34 / PR5 PRs
Running now6 sessions · 5 machines
- payments-core · Claude Code · mac-7f3arunning commands · services/billing$4.20
- data-pipeline · Claude Code · lin-2c91loop · Bash ×14$11.80
- web-client · Claude Code · mac-5d3ewriting code · src/checkout$2.65
- admin-panel · Codex · win-9b10waiting on a person · 6 min$1.10
- infra · OpenCode · lin-4a77waiting on a person · 14 min$0.90
- payments-core · Claude Code · mac-7f3awriting code · services/ledger$3.05
Example figures — one team of eight, one week.
Three questions
Three questions a CTO cannot answer today.
What did last quarter's agent spend buy?
Cost per merged pull request, by repository and release, whichever agent ran it.
“doesn't allocate tokens or cost to individual commits or pull requests”
DX, AI Code Insights each daythe invoice: one number, $7,409
One developer’s Mac, 29 September to 28 October 2026, at list price. Which agent is burning budget right now?
Every session live on every machine, each plan's distance to its limit, the looping one marked.
“lists only sessions the desktop app runs itself”
Claude Desktop docs One developer’s Mac, 13 to 21 October 2026: one Claude plan’s seven-day window, read every hour. How much went through Bedrock, seats and personal plans together?
Each session priced the same way, whichever route paid for it.
“Claude Code does not send metrics from your cloud back to Anthropic”
Anthropic, Manage costs Today
3 dashboards
two of them blind
With Caprock
one total
Bedrock + seats + Max · list price
The charts are measured, one developer’s Mac. The third row: example figures, one team of eight. Quotes read 27 October 2026.
How it works
A developer installs it. Your server adds it up.
My sessionstoday
- payments-core · Claude Coderunning commands$4.20
- payments-core · Claude Codewriting code$3.05
- infra · Codexended 12 min ago$1.40
What the laptop sendsper session, per hour
- repo
- payments-core
- pr
- #482
- release
- v1.9.2
- tokens
- 84,120
- cost_usd
- 4.20
- machine
- mac-7f3a
prompt · code · commandsno field
Running now
6
on 5 machines · 1 loop
By repositoryper merged PR
- payments-core$35 / PR
- web-client$29 / PR
- data-pipeline$41 / PR
Example figures — one team of eight, one week.
What you see
Live, per pull request, and who needs a person.
Live
Every session on every machine, as it runs.
- payments-core · Claude Code · mac-7f3arunning commands · services/billing$4.20
- data-pipeline · Claude Code · lin-2c91loop · Bash ×14$11.80
- web-client · Claude Code · mac-5d3ewriting code · src/checkout$2.65
Per pull request
What each merged pull request cost, whichever agent ran it.
Merged
58
GitHub
Per PR
≈$33
estimate
Unlinked
18%
$420
- #482 Retry failed payouts$46.20payments-core · 3 sessions · Claude Codeby commit
- #479 Apple Pay in checkout$38.10web-client · 2 sessions · Claude Code, Codexby commit
- 56 more merged PRs$1,805.70
Who needs a person
Looping, or waiting on a person, with the cost so far.
- data-pipeline · Claude Code · lin-2c91loop · Bash ×14$11.80
- admin-panel · Codex · win-9b10waiting on a person · 6 min$1.10
- infra · OpenCode · lin-4a77waiting on a person · 14 min$0.90
Example figures — one team of eight.
Already using something?
Each one sees its own slice.
- Vendor dashboards
- Anthropic and Cursor meter their own agent on their own billing route.
- Cockpits
- Orca, Conductor and T3 Code list the sessions they start.
- DX, Jellyfish, LinearB
- Read vendor APIs and report per person.
- Caprock
- Every session on every machine, whoever started it, priced however you pay.
Orca vs Caprock
Where it is ahead. A code editor and a browser inside the app, native phone apps, any CLI agent free, a far larger community.
What Caprock adds. The sessions Orca never started, and what each pull request cost.
Read the comparison →
T3 Code vs Caprock
Where it is ahead. A server you host yourself, several machines in one app, native phone apps with push alerts.
What Caprock adds. One ledger across every agent and machine, whoever started the session.
Read the comparison →
Herdr vs cmux
Where it is ahead. A terminal runtime or a Mac terminal app: both run agents well and neither keeps a ledger.
What Caprock adds. Where each one fits, and what Caprock adds beside it.
Read the comparison →
Each as its own documentation describes it, read 27 October 2026.
Trust
Counted by work. Kept by you.
By repository, PR and release
A per-person view is a documented switch, never the default.
Every developer sees the same figures
- spend
- $2,310
- merged PRs
- 58
- per merged PR
- ≈$33
- spend
- $2,310
- merged PRs
- 58
- per merged PR
- ≈$33
The same aggregates the lead sees, and what leaves the laptop.
No field for a prompt
Sent to your server
- repo
- payments-core
- pr
- #482
- release
- v1.9.2
- tokens
- 84,120
No field for it
- prompt
- reply
- code
- command output
Your server gets the repository, pull request, release and cost. A prompt, a reply or a command's output has no field to travel in.
Your server, your network
✕ caprock.dev · nothing sent
One binary behind your VPN or SSO. Nothing is sent to Caprock.
Price
From $500 a month, billed yearly.
One flat price for the whole team. Community Edition stays free.Twenty minutes on your stack, four short questions, then a slot.
Before you book
Five questions.
Is it a bill?+
No. Every dollar Caprock shows is an estimate at API list price, as Claude Code's own /usage is: not a bill, and on a flat plan not money saved. In the app, a plan's line also shows the plan dollars and the share of the weekly limit.
Does it see everything?+
No. Sessions that run in a vendor's cloud or inside CI are not observed; Caprock sees the machines it is installed on.
Can my developers see what leaves?+
Yes. Settings → Privacy on each laptop lists what leaves it, and every developer sees the same figures the lead sees.
Per person?+
A documented switch, off by default. When a team turns it on, everyone on the team sees the same screen, the admin's figures included.
Where does it run?+
On your server, in your network, as one binary. Laptops push sums to it; nothing pulls from a laptop, and nothing is sent to Caprock.