Meters every major AI coding tool your team already runs.
Claude Code
Cursor
Windsurf
Antigravity
4 tools
one shared dashboard
< 5 min
to the first synced session
0 lines
of your code ever read
$0.0001
cost resolution per session
AI changed how we write code. Nobody knows what it’s costing.
Every team is already spending on AI. On an invoice, every interaction looks identical, so no one can tell which developers, repositories, or workflows created any value.
Cloud got its observability layer a decade ago. AI coding doesn’t have one yet.
01
Cost
Where is the spend actually going?
There is no attribution across developers, teams, or repositories. Finance sees one number at the end of the month that keeps climbing, and nobody can break it apart.
02
Efficiency
Are teams getting leverage from AI?
No visibility into wasted tokens, idle seats, oversized models on trivial work, or the sessions that grind for two hours and ship nothing. None of it surfaces until someone measures it.
03
Impact
Is AI improving engineering outcomes?
Nothing connects AI usage to shipped work. The one number that would settle the argument, what a shipped change actually costs, does not exist anywhere.
This is not a niche line item. It is the fastest-growing budget in most engineering organizations.
84%
of developers use or plan to use AI coding tools
Stack Overflow Developer Survey 2025
91%
of organizations run more than one AI coding tool
GitLab Global DevSecOps Report 2026
51%
of professional developers use AI every day
Stack Overflow Developer Survey 2025
PlanckSpace makes AI coding measurable.
A lightweight local agent continuously attributes AI coding activity across developers, repositories, and teams, turning raw usage into engineering decisions you can defend in a budget meeting.
01
Measure
Every AI session, token, and dollar attributed to the developer, team, repository, and model that generated it.
attribution, not invoices
02
Optimize
Concrete actions, priced: reclaim idle seats, route the right model, shrink the context you re-send, end the sessions that never converge.
a ranked queue, in dollars
03
Connect to outcomes
Usage tied to what actually shipped: repository activity, developer throughput, and the cost of every session that made it to main.
cost per shipped session
04
Verify
Every recommendation re-measured from your own telemetry after the fix, so savings are booked from data instead of claimed from a model.
a ledger, not an estimate
Never your code. Never your prompts.
The product
Turn AI coding into clear engineering decisions.
Every session your team runs becomes attributed, explainable spend, tied to the work that actually shipped. In the dashboard, in the CLI, and in your editor.
One number the whole company agrees on.
Live workspace spend across every tool, model, and developer, with cost per shipped session sitting next to it. Engineering, management, and finance read the same figure at the same time.
updates as sessions close
split by team, repo, or developer
reconciled against the provider invoice
Usage value
$1,134↑ 120%
Cost / shipped
$25.21↑ 458%
Shipped rate
88%↓ 8.4 pts
Spend over time
$1,132↑ 120%
TotalBy teamBy repo
Week 1Week 2Week 3Week 4
Waste, itemized in dollars
Idle seats, cache misses, and oversized models on trivial work arrive as line items you can act on, each priced.
Ranked by impact$594/mo
Context re-reads, api-service3,592× cache to input$204
Marathon sessions that never shipped3 sessions, 152 turns each$195
Opus on trivial editsroute to Haiku$104
The invoice, reconciled
Attributed usage is matched against what your providers actually billed, and the difference is explained rather than rounded away.
October, reconciled
Invoiced by providers
$6,120
Attributed by PlanckSpace
$5,780
Unattributed
$340
One engineer is running without the agent installed. Invite them and the gap closes.
Alerts before the invoice
Budgets track pace through the month and anomalies fire the day they happen, so nothing arrives as a surprise in accounting.
October budgeton pace
$5,120of $8,000
projected close $7,430
Session outliercritical
$190 against a $16 median, posted to #eng-spend
Insight in your editorhigh confidence
CLAUDE.md is re-read 212× per session in api-service$34/mo
Fix nowFix with Claude Codebacked up, one click to undo
Booked from telemetry
$212saved since May 12
Context split, api-service$28/mo verified Jun 02
Model routing, workspace$19/mo verified Jun 21
Fix it in the editor. Book the saving only when it's real.
The extension for VS Code, Cursor, and Windsurf applies mechanical fixes in place and hands the judgement calls to your own Claude Code with the measured data attached. Then it watches your telemetry and books what actually landed.
every fix backed up and reversible
re-measured across the sessions that follow
unconfirmed savings are never counted
Per-team attribution
Spend rolls up by team, repo, and developer, with the unattributed remainder always visible.
Session-level receipts
Every dollar traces back to one session: model, token mix, cache efficiency, and repo.
Finance-ready exports
Chargeback and showback CSVs cut by team or cost centre, sized for the monthly close.
Six detectors. Every one shows its work.
Each targets a specific way AI spend leaks, and each carries the measurement that found it: the signal, the threshold, the arithmetic, and the re-measurement that confirms the fix worked. Nothing here needs to read your code.
01
Context re-read metering
high confidence
Finds the stable context you pay to re-send every single session.
detector context-bloat
Measured signal
3,592×
cache reads per input token
Cache-read ÷ input tokens
3,592×
Fires above
0.40×
Avg input / session
48,210 tokens
Sessions measured
51 in 30 days
How the number is built
$1,134 window spend × 18% recoverable × (30 ÷ 30)
= $204.12 / mo
The fix
Move CLAUDE.md's topic sections into .claude/docs/ and leave a pointer index in the root. Each file then loads only in the sessions where its topic comes up. Not @import: imports load in full at every conversation start, so they move text around without saving a token.
Read to know thiscacheReadTokensinputTokenscostUsdstartedAtcounters only, no code, no prompts
02
Premium-model routing
high confidence
Catches the expensive model doing work that never shipped.
detector model-routing
Measured signal
62%
of sessions on Opus, 24% of what shipped
Opus share of sessions
62.0%
Opus share of shipped work
24.0%
Fires when
session share > shipped share
Spend on Opus that didn't ship
$104.30
How the number is built
$104.30 non-shipping premium spend × (30 ÷ 30)
= $104.30 / mo
The fix
Add a routing rule to CLAUDE.md: default exploration and iteration to a Sonnet-class model, escalate to Opus for architecture, tricky debugging, and large refactors. The gap between those two percentages is the money.
Read to know thismodeloutcomecostUsdgitAuthorEmailcounters only, no code, no prompts
03
Cache-hit recovery
high confidence
Spots high token volume paying full input price on every turn.
detector cache-efficiency
Measured signal
11.4%
cache-hit rate across 8.2M tokens
Cache-hit rate
11.4%
Fires below
30.0%
Token volume
8.2M
Volume floor
1.0M
How the number is built
7.3M input × 50% cacheable × 90% saved × $3.00/M
= $98.55 / mo
The fix
Put the stable prefix (project overview, conventions, standing guidance) first and unchanged, ahead of anything task-specific. The cache keys on that prefix, so anything dynamic above it invalidates everything below.
Read to know thisinputTokenscacheReadTokenssessionIdcounters only, no code, no prompts
04
Marathon-session detection
high confidence
Flags long sessions that grind past the point of converging.
detector marathon-sessions
Measured signal
6 sessions
averaging 152 turns, none shipped
Marathon sessions
6
Turn threshold
≥ 30 turns
Avg turns
152
Combined cost
$195.40
How the number is built
$195.40 across 6 non-shipping marathons × (30 ÷ 30)
= $195.40 / mo
The fix
Past roughly 30 turns the accumulated context is re-sent on every turn and the session rarely converges. When one stalls, start fresh with a tighter prompt carrying what you learned. Restarting is cheaper than pushing.
Read to know thisturnCountoutcomecostUsdcounters only, no code, no prompts
05
Idle-seat reclaim
high confidence
Prices the seats nobody is using against what you're billed.
detector seat-efficiency
Measured signal
4 seats
consuming under 10% of their cost
Dormant seats
4
Dormant below
10% of seat cost
Seat cost
$30.00 / mo
Avg consumption
$1.14 / mo
How the number is built
4 dormant seats × $30.00 seat cost
= $120.00 / mo
The fix
Reclaim or reassign at the provider. This one is deliberately never automated. It is a billing change against a real person's access, so PlanckSpace surfaces it and hands you the list, nothing more.
Read to know thisgitAuthorEmailcostUsdtoolseatCostUsdMonthlycounters only, no code, no prompts
06
Telemetry-verified savings
high confidence
Re-measures after the fix, so savings are booked and never claimed.
detector insight-verification
Measured signal
94%
of the measured problem eliminated
Baseline at detection
3,592×
Re-measured after fix
212×
Confirms above
30% improvement
Post-fix sessions needed
≥ 3
How the number is built
$204.12 estimate × 94% actually realised
= $191.87 / mo booked
The fix
The baseline is snapshotted at first detection, then re-measured over the sessions that ran after the fix. A fix you marked done whose metric never moved stays 'claimed' and is never counted. A metric that improves on its own still verifies.
Read to know thiscacheReadTokensinputTokensstartedAtstatusChangedAtcounters only, no code, no prompts
Three commands, and the whole team is measured.
Developers keep the tools they already use. Nothing proxies your traffic, nothing enters your build, and there are no API keys to rotate. Once we have provisioned the workspace, we are on the call for all three steps.
01
Install the CLI
One command. The installer sets up the planck binary and a background daemon that watches local session logs.
$ curl -fsSL https://planckspace.dev/install | sh
✓ planck v1 installed
02
Connect your workspace
Log in once and every future session syncs automatically: token counts, model, cost. Metadata only.
$ planck login
✓ workspace linked: acme-eng
daemon syncing, 12 sessions found
03
Invite the team
Teammates join with an email invite. Spend appears in the shared dashboard, and in the VS Code extension right inside the editor.
$ planck status
workspace acme-eng, 11 members
✓ live at console.planckspace.dev
Privacy by architecture
Never your code. Never your prompts.
PlanckSpace meters receipts, not work. It parses the metadata your AI tools already write to disk: token counts, models, costs. Your source, prompts, and responses never leave the machine. Developers can verify exactly what syncs with planck inspect.
Sync is opt-in per machine. Revoke a device at any time.
Synced
Model and token counts
Session cost and duration
Tool and editor used
Repo name and git author
Never synced
Source code
Prompts and responses
File contents
Anything you typed
The same numbers, three very different questions answered.
For leadership
Defend the AI budget with numbers.
+One figure for total AI spend, reconciled against provider invoices
+Cost per shipped session: the ROI number, by team
+Savings verified from telemetry, not a vendor's estimate
+Board-ready exports without asking engineering
For managers
Know where the budget actually goes.
+Per-team and per-repo attribution, updated live
+Budgets with pace alerts before the overrun, not after
+Waste surfaced as line items with a dollar value
+A ranked queue of fixes, highest recovery first
For developers
Your usage, not your keystrokes.
+Personal cost and cache-efficiency insights in the editor
+One-click fixes, backed up and undoable
+Metadata only, verifiable with planck inspect
+No screenshots, no timers, no surveillance
Fair questions, straight answers.
Anything we have not covered? Ask us directly. A human replies.
Stop estimating. Start measuring.
Book 30 minutes. We’ll show you what your team’s AI coding costs, what it’s returning, and then set the workspace up with you.