Quick answer
The local whoburnedmore CLI command is npx whoburnedmore --local. Decline its optional publish offer to keep the report local. The standard command signs in and submits daily aggregates. Neither mode submits prompts, source code, or filenames.
What this guide answers
The reader needs a team policy for AI coding tools.
Best for
Engineering managers, platform teams, and startup operators.
Key facts, checked November 8, 2026
- Claude Team and Enterprise admins can turn on usage credits and then set spend limits at the organization, group or member level. Console (API) users get workspace spend limits and rate limits. source
- GitHub Copilot budgets can be set per enterprise, organization, cost center and user. Copilot paid usage is on by default, and admins can turn it off with the 'AI credits paid usage' policy. source
- Cursor Teams supports one team-wide spend limit, fixed or per seat. Enterprise adds member and group limits, and enforcement can lag briefly. source
- DORA cites Shopify's 'circuit breakers', which cut access when someone's daily spend spikes, as a model guardrail. It warns that individual token scores are harmful. source
Start with approved tools
Tool sprawl makes cost invisible. List the default tools and the cases where developers can use alternatives.
- Default IDE tool
- Default terminal tool
- Prototype tool
- Exception process
Use routing rules
Routing rules reduce duplicated spend. They also make it easier for developers to pick the right tool without asking every time.
- Small edit
- Large refactor
- Prototype
- Code review
- Incident
Prefer soft controls
Strict blocks can slow valuable work and encourage shadow tools. Soft warnings and reviews usually work better unless abuse is repeated.
- Warning
- Review
- Exception
- Block
Protect privacy
Usage tracking should not require shipping source code or sensitive prompts to a leaderboard. Define what is collected and what stays local.
- No source code.
- Aggregate usage.
- Opt-in public boards.
Map each policy line to a real control
A policy only holds if a vendor setting enforces it. Before you publish one, check that every soft or hard limit you write down has a matching switch. Claude: group and member spend limits on Team and Enterprise, or workspace limits on the Console. Copilot: user and cost-center budgets, with 'Stop usage when budget limit is reached' for a hard stop. Cursor: a team or per-seat spend limit, with per-member limits only on Enterprise. If a control doesn't exist, the line is a guideline. Label it as one.
- Soft limit: vendor alerts, such as spend alerts and budget notifications
- Hard limit: a budget with a hard stop, or a member spend limit
- Circuit breaker: a daily spike check that pauses access pending review
Clauses worth copying
These five clauses cover the usual sources of argument. First, approved tools: one default IDE assistant and one default terminal agent, plus anything else under a 30-day trial with a named owner. Second, an exception path: incidents and migrations may exceed caps, with a one-line note within 48 hours. Third, privacy: no tool may upload prompts, source code or raw logs for usage tracking, only aggregates. Fourth, measurement: usage is reviewed at team level, never ranked for performance reviews. Fifth, review: a 15-minute weekly look at the top three burns and one rule change.
Measuring before you set caps
Set caps from two to four weeks of real data, not guesses. For local coding agents, npx whoburnedmore --local builds an offline dashboard from 22 supported tools and sends nothing. That suits a private baseline. When the team is ready to compare, an organization board lets members join with npx whoburnedmore --org=SLUG --pass=CODE. Members share daily token and cost totals plus usage metadata such as tool-call and skill counts, never prompts or code.
Common questions
Should an AI spend policy use hard limits?
Use hard limits as a backstop for runaway spend, and soft alerts for everyday guidance. Hard caps hit during incidents push people toward unapproved tools.
Can I set per-developer AI spend limits?
Yes, in Claude Team or Enterprise (member-level limits), Copilot (user-level budgets) and Cursor Enterprise (member limits). Cursor Teams supports a per-seat team limit.
What should a privacy clause say?
That trackers may collect only aggregates, such as date, tool, model, token counts, estimated cost and tool-call counts. Never prompts, code, file paths or raw logs.