burn-baby-burn: a token bonfire for the metrics era, and a joke with teeth

The premise, stated without shame

The repository's subtitle is the whole argument: because nothing gets you promoted faster than a six-figure token bill. The tool itself is a bash command that burns Claude Code or Codex tokens on purpose:

burn 50000  # burn 50,000 tokens

The joke lands because the target is real. Organizations now measure engineering by AI usage dashboards, internal leaderboards rank teams by tokens consumed, and vendor pitches quote those curves as productivity. Once consumption is the metric, consumption is what gets optimized, and a tool that generates usage with zero output is simply the reductio.

What makes the repository more than a tweet is that the tool works. It is not a mockup; it drives an authenticated agent and spends real quota, which is the difference between joking about Goodhart's law and demonstrating it on your own invoice.

Install, usage, and the wink in the flags

Installation is the standard pipe-to-shell, dropping a single command into the user's local bin without sudo or a package tap, with the prerequisites stated plainly: Claude Code or Codex authenticated on the path, plus a JSON processor available from the usual package managers.

curl -fsSL https://raw.githubusercontent.com/dtnewman/burn-baby-burn/main/install.sh | bash

The usage block is where the satire sharpens into commentary:

burn 10000                            # the minimum
burn 50000 --model haiku              # cheap and fast
burn 100000 --model sonnet            # walk away

Read the flags as a genre portrait. A minimum burn, for the cautious careerist. A cheap, fast model, for padding stats at low cost, which is exactly the optimization a metrics-gaming team would make. A larger burn on the stronger model, captioned as walk away. And the documented fourth form switches the backend to burn the other vendor's tokens instead, because loyalty is not part of the joke.

The command syntax is doing parody and documentation simultaneously, and the screenshots deliver the punchline: a chart of tokens consumed, rising beautifully, signifying nothing.

Features, reviews, and the enterprise edition

The features list is four lines of pure indictment: make the CEO see how productive you are; investors will see how AI-innovative your company is; top the internal leaderboard; pad your OKRs. Every line names a real incentive that exists in organizations right now, which is why the joke stings rather than sprawls.

The reviews section supplies fabricated testimonials with attributed sources: a CEO marveling at five-times productivity a week after installation, a venture firm doubling an offer after seeing usage stats in a pitch deck, and the user themselves, months later, enjoying the promotion. The quotes are fake; the dynamics they describe are observable in any organization that started measuring AI usage before deciding what it meant.

Then the enterprise edition: a paid version in development with four-times burn rate, single sign-on support, and role-based burn access. It is a perfect three-beat parody of enterprise software roadmaps, and the contact-us line at the end is the part that should make you check whether you have ever written one sincerely.

The licence note, and the honesty underneath

The licence line closes the document with two jokes and one confession: MIT, but it is all vibe-coded so who really knows who owns it, followed by the instruction to burn responsibly. The confession inside the joke is the current moment in miniature, an era of tools whose provenance and ownership are genuinely murkier than their licenses claim, said quickly enough to be funny.

Underneath the comedy, the repository is competently put together: a single-purpose script, sane flag surface, real backends, documented prerequisites, an installer, and a chart. The craft is the delivery mechanism for the satire, and the satire would not land if the tool did not work.

There is even a legitimate use hiding in the joke: burning a known quantity of tokens is a crude but honest way to test quota accounting, spend alerts and dashboards before they meet real workloads. A tool built as a punchline accidentally doubles as a metering test, which is more than most satire manages.

Who installs a joke, and what sits beside it

As software, almost nobody should run this with intent: its entire function is spending money to decorate a chart. As commentary, its audience is everyone who administers, reads or appears on an AI-usage leaderboard, and the correct use is reading it, wincing, and then asking what your own dashboards reward.

The neighbours are the things it mocks. Usage dashboards and internal leaderboards present consumption as productivity with no satirical intent. Budget-justification decks quote token growth as adoption. And on the sincere-tool side, agent harnesses with verbose logging modes produce much the same effect at one remove, burning context and quota for visibility rather than output.

The last push was on 2026-06-09, there are no releases, and the repository is two scripts and some images. Judged as a tool, it is a lighter; judged as writing, it is one of the more efficient pieces of organizational satire you will find in a licence file's neighbourhood, and it costs nothing to read before it costs you tokens to run.

Editorial conclusion

burn-baby-burn fits readers who administer or appear on AI-usage metrics and want the sharpest available one-command rebuttal to consumption-as-productivity thinking, and engineers who want a crude quota-metering test with a sense of humour. It does not fit anyone's actual workflow, since its function is deliberate waste, or organizations without the self-awareness to enjoy the enterprise-edition joke. Verify first: nothing, which is the point of a joke repository; if you do run it, a spend limit on the account you point it at, and a small model flag. The licence is MIT, with the ownership caveat the author supplies themselves, and the last push was on 2026-06-09.

Frequently asked questions

Official sources

1. dtnewman/burn-baby-burn on GitHub
2. Issues
3. License: MIT
4. README