Feature/Cost and budgets

Know what every run would cost, and spend less of it.

Nautelle records the tokens of every model request your agents make and prices them, so you see spend per project, model, issue, slot and card. Budgets alert or stop new work, and Token Saver keeps noise out of what agents read.

Read the guide Self-host it

Usage: API-equivalent cost over time, budget status and a breakdown by project. Demo project, fictional data.

API-equivalent cost, not a bill

Every dollar figure is what the recorded tokens would cost at public API prices, even when your agents run on a subscription. It is the one number you can compare across runtimes, models and months. Nautelle bills nothing.

Tokens arrive live from OpenTelemetry on loopback, and transcripts fill in the rest
Prices live in versioned tables; old versions are never changed in place
A model with no price shows as unpriced, never as $0
Only metadata is stored: no prompt, response or tool text

C-44 · round 2$1.84

cacheRead
input
output
cacheWrite5m
cacheWrite1h
reasoning ⊂ output

A model with no price shows as unpriced, never as $0.

Down to the turn that burned the tokens

Sessions lists every agent session you can see, with turns, requests, tokens, cost and duration. A session’s page is a tree with a waterfall.

Turns, model requests, tool calls by name and duration, and subagent sessions
Time bars and totals on every node
Breakdowns by project, model, runtime, issue or slot on the Usage page
Codex sessions are listed but carry no tokens or cost yet

A session tree: turns, model requests, tool calls and subagents with tokens and cost. — A session tree: turns, model requests, tool calls and subagents.

Budgets that alert, or stop new spend

A budget caps the API-equivalent spend of a project or a user per day, week or month, in its own time zone. Thresholds notify once per period, in-app and on Telegram.

Thresholds default to 50, 80 and 100 percent
With stop, new runs and orchestrator starts are refused; running work keeps going
An admin can override a stop for the rest of the period, with a reason
Every change and override is audited

northwind-web · monthwhen reached: stop

$444 of $600

50%80%100%

50% notified once this period
At 100%, new runs and starts get budget_exceeded

Token Saver: agents read less noise

Token Saver changes what agents read, never what people read. Covered commands are filtered to the lines that matter, only the MCP servers you allow load, and whole reads of very large files are guarded.

On by default, per project
Errors, failures, warnings and stack frames are never removed
The full output stays on the runner machine; NAUTELLE_RAW=1 skips the filter once
Savings are estimates, shown with a tilde, per card and per filter

$ bun test
✗ invoices › totals › rounds to cents
  expected 1999, received 2000
  at totals.test.ts:41:7
411 pass · 1 fail

[token saver] 1,284 lines hidden · full output kept on the runner

saved ≈ $0.21· exit code unchanged · errors kept

See what it kept out

Savings show on the card next to its cost, on the Usage page with cost per run before and after, and on the project’s own Token Saver dashboard by time, filter and run kind.