seatledger · Claude Code and Codex
What your coding-agent seats and tokens bought, per developer, model and client — and the day the rate changed.
$ npx seatledgerIt reads the transcripts Claude Code and Codex already keep on your machine and prints tokens and API-equivalent dollars by day, model and client, your burn toward limits you set, and every step change in how requests use tokens. MIT, no account, no telemetry; it sends nothing unless you create or join a team ledger. Source · npm
What it prints
$ npx seatledger
seatledger — what your coding-agent seats and tokens bought, from this machine
Today, Wed 7 Oct
requests input output cache write cache read tokens API-equiv $
Claude Code · claude-opus-5-5 122 1.1K 115K 5.29M 5.37M 10.8M $29.81
Codex · gpt-5.6-sol 28 98K 23K 46K 882K 1.05M $1.43
Claude Code · claude-haiku-4-5-20251001 9 8.1K 360 0 0 8.5K $0.01
total 159 107K 138K 5.33M 6.25M 11.8M $31.25
Last 7 days, Thu 1 Oct – Wed 7 Oct
requests input output cache write cache read tokens API-equiv $
Thu 1 Oct 100 119K 83K 2.17M 3.66M 6.03M $13.88
Fri 2 Oct 118 109K 113K 2.57M 5.53M 8.32M $16.80
Sat 3 Oct 48 3.1K 44K 1.54M 2.52M 4.10M $9.07
Sun 4 Oct 39 3.0K 33K 1.37M 1.35M 2.75M $7.75
Mon 5 Oct 110 188K 96K 2.58M 4.11M 6.98M $16.70
Tue 6 Oct 140 107K 127K 3.60M 5.48M 9.32M $22.21
Wed 7 Oct 159 107K 138K 5.33M 6.25M 11.8M $31.25
total 714 637K 634K 19.1M 28.9M 49.3M $117.65
Limits: none set. `seatledger limits set --window 5h --tokens N` records your own;
seatledger does not guess a vendor's plan limits.
Vendor readings, as the client last wrote them to disk
Codex 5-hour window: 26% used, resets 18:00, plan plus (read 13:43)
Codex weekly window: 7% used, resets Sun 11 Oct 18:00, plan plus (read 13:43)
API-equiv $ = what these tokens would cost at the vendor's API list prices. Not what you were
billed; a seat is a flat fee with an allowance.
Read 180 Claude Code transcripts and 32 Codex transcripts, on this machine only. Nothing was sent
anywhere.
More: seatledger report --by project · seatledger rates · seatledger limits set --window 5h --tokens NSynthetic history: the real output of npx seatledger demo, which writes generated transcripts in the real formats and reads them. None of it is anyone's usage (as of 2026-10-07).
When the rate changes
npx seatledger rates looks, per client and model, for a day where the same work started using a different number of tokens and stayed there: tokens per request, the cache read less and written more, more requests per prompt, or Codex's own 5-hour window filling faster for the same tokens. From the synthetic history:
From 22 Sep (Codex 0.159.0, the first day on it), tokens per 1% of the 5-hour window halved across Codex models — consistent with a smaller allowance per token: the same tokens now use more of the 5-hour window (usage on other devices on the same account also lowers this). It starts with Codex 0.159.0 (previously 0.158.0).
It says what a step is consistent with, never what caused it: it sees the requests on your machine, not the vendor's servers. A change in your own work moves these numbers too, and when no client update coincides, it says so.
The team ledger
For the lead who pays for the team's seats: every developer on one page, per model and client, kept longer than a laptop keeps transcripts, and an email the first time a step change shows up on anyone's machine.
- up to 3 developers
- 30 days of history
- rate changes on the dashboard
- up to 10 developers
- 13 months of history
- rate changes by email and Slack
- CSV export
- up to 50 developers
- 13 months of history
- rate changes by email and Slack
- CSV export
Signing in with GitHub creates the team and its invite link. From a terminal instead: npx seatledger team create creates it and prints the invite link. Each developer then runs npx seatledger team join <link> once and npx seatledger push after. Paid plans are billed monthly by card through Stripe and cancelled in its portal.
Seeing which people hold paid coding seats at more than one vendor, and premium seats that never ran over, from your admin exports, is not built. I want this — the click is counted; nothing else is sent.
The free alternative: Ramp
Ramp's AI Token Spend Management, launched 16 July 2026, is free and needs no Ramp card. With your admin API keys it reads the cost and usage APIs of Anthropic, OpenAI, Gemini and Cursor (its help centre adds AWS Bedrock and Fireworks AI), and breaks spend down by provider, model, user, API key and team, with a weekly briefing and anomaly alerts; spend limits are paid. If what you want is the company's whole AI bill, including the API keys your servers use, start there: seatledger reads no admin API, sees only Claude Code and Codex, and does not see that spend.
What seatledger adds
- Local session logs. It reads the transcripts Claude Code and Codex keep on each developer's machine, so it needs no admin key, and one developer can run it alone.
- Usage on subscription seats. Claude Pro, Max and Team seats and Codex on a ChatGPT plan, in tokens and API-equivalent dollars. Ramp imports subscription usage only from Claude Enterprise and ChatGPT Enterprise; its help centre says Claude Team, ChatGPT Business and individual plans are not supported.
- Rate changes. The day the same work started using more tokens per request, or the cache less, with the client version where it starts. Ramp's alerts watch spend; a spend total does not show tokens per request.
From Ramp's launch post (16 July 2026), Ramp's product page and Ramp's help-centre setup article, read .
Exactly what a push sends
One JSON document, seatledger.push/v1, with these fields and no others; the service refuses any field it does not know and any text that could hold a sentence. npx seatledger push --dry-run prints the document and sends nothing.
The document
- schema
- the constant seatledger.push/v1
- days
- every local day the push covers; the stored rows for those days become exactly rows
- rows
- one per day × client × client version × model, with the fields below
- findings
- what seatledger rates found, as numbers, with the fields below
Each row
- day
- the developer's local calendar day
- client
- claude-code or codex
- client_version
- the version the client wrote, e.g. 2.1.286
- model
- the model id the client wrote, e.g. claude-opus-5-5
- project
- the working-directory folder name, only with --project-names
- requests
- how many requests
- input_tokens
- a count
- output_tokens
- a count
- cache_read_tokens
- a count
- cache_write_tokens
- a count
- api_equivalent_usd
- those tokens at the vendor's API list prices (null if unpriced)
Each finding
- client
- claude-code or codex
- model
- the model id, or * for Codex's own window readings
- date
- the day the step starts
- version
- the client version in use from that day
- previous_version
- the version before it
- version_changed
- whether that day was the first on the version
- changes
- per metric: its id, the median before and after, and the change
- cache_reads_per_request
- before and after
- cache_writes_per_request
- before and after
Never sent
- prompts, responses, tool output or file contents
- file paths, session ids or request ids
- project folder names, unless a developer pushes with --project-names
See the privacy note for who sees what and how long it is kept.
Guides
For agents
POST /api/seatledger/v1/teams {"accept_terms": true} creates a free team with no human and returns an owner key and an invite link; POST /api/seatledger/v1/checkout returns a Stripe link a person can pay. The CLI's MCP server reads the team with seatledger_team_summary. Machine-readable: llms.txt, pricing.json, openapi.json.