If you've ever stressed about watching your Claude or Gemini API tokens tick away unused before monthly reset, schwabe has heard your cry. This newly surfaced npm package—installable via npm i -g schwabe (alias: tokenburner)—is a terminal-based token burner that spins up fleets of real AI agents to absolutely obliterate whatever quota you've paid for, logging every single token spent in the process.

The Schwabe Philosophy

The project name comes from "Schwabe," referring to people from Baden-Württemberg, Germany—stereotypically known for not wasting a single Pfennig. As the README puts it: 'Paid for it? Then you use all of it, bis auf den letzten Tropfen.' The package is built around that exact mindset applied to AI token limits. You paid for 100,000 tokens? They're going to get used. All of them. Every single one logged to ashes.jsonl and burns.csv with lifetime totals persisting across runs.

Burn Modes and Fleet Architecture

The tool runs entirely in the terminal with a btop-style live dashboard—heat-graded progress bars, per-agent spinners, tokens/sec graphs, and running LEDGER totals that auto-adapt to light or dark themes. Users pick from three budget tiers via flags: --haiku (cheap), --sonnet (balanced), or --rich/opus (expensive). The --count N flag spawns a parallel fleet of agents, with --parallel controlling concurrency at roughly 1 GB RAM per agent. Each AI worker gets assigned absurd tasks—'a sonnet about a runaway semicolon,' 'an argument over whether water is wet,' release notes for the heat death of the universe—to ensure maximum token consumption.

The --Forest Offset and Other Features

For environmentally-conscious Schwabes, --forest plants ASCII trees into FOREST.txt while calculating virtual CO₂ savings against real per-species sequestration rates. This is labeled 'Certified carbon-negative™' with a parenthetical '(It is not.)' The --dry flag runs a mock backend for safe previews with zero spend. A share.js script reconstructs your burn session and generates prefilled social links—behind a confirmation prompt, never auto-opening browsers. Everything is built clean, dependency-free, ESM with a registry pattern in lib/core/registry.js making it trivial to add new backends or share platforms.

Claude Code Integration

Inside Claude Code itself, the /burn [N] command runs parallel agent swarms and forges the funniest wreckage into HALL_OF_FLAME.md. This isn't just a joke tool—it has real utility as a stress-tester or chaos engineering playground for AI workflows. The backends currently supported: claude, gemini, codex, and mock (for testing). Everything is unit-tested with Node's built-in test runner with zero test dependencies.

Key Takeaways

  • Installs globally via npm i -g schwabe, runs entirely in terminal
  • Supports Claude, Gemini, Codex backends plus a mock mode for safe previews
  • Fleet-based parallelism with configurable concurrency (~1 GB RAM per agent)
  • Logs all activity to burns.csv (metrics) and ashes.jsonl (output), persists across runs
  • Includes --forest option for satirical carbon offset with ASCII trees

The Bottom Line

Token expiration anxiety is real, and schwabe weaponizes that pain into a terminal experience designed for maximum quota destruction. It's the kind of project that makes you laugh, then realize you've absolutely been there—staring at unused tokens with guilt as the clock runs out. WTFPL license means you can do whatever the fire wants with it.