Open source · Research · MIT

Stop wasting
tokens.

A team of independent researchers obsessed with saving tokens. We build open-source tools that make AI agents cheaper and better.

Try VBW Join the movement

The method

Agents spend tokens you never see.
We make the spend visible, then smaller.

Stop

Agents re-read, retry and rework without telling you. Step one is refusing to pay for it blind.

Count

Measure before you claim. We publish the method before we publish the number.

Cut

Fewer tokens, same work, proven by tests instead of vibes.

Flagship · by SWT Labs

VBW

Vibe Better With Claude Code

Claude says it's done. VBW makes it prove it. A plugin that turns Claude Code into an engineering team kept honest by a deterministic kernel.

Acceptance tests are written first, and fail first.
The contract is sealed with SHA-256, so it cannot quietly change.
No AI in the rule-keeper: probabilistic in, deterministic out.

Enter VBW Read the source ↗

Install in two lines/plugin marketplace add swt-labs/vibe-better-with-claude-code-vbw/plugin install vbw@vbw-marketplaceThen run/vbw:vibe

★ 8414 forksMIT

The receipts

Numbers, with their
method attached.

Spawn and context, per phaseVBW 2: 72% less

Stock 87,100

VBW 1 26,535

VBW 2 24,000

All coordination, one phaseVBW 2: 72% less

Stock 97,900

VBW 1 29,885

VBW 2 27,000

All coordination, five phasesVBW 2: 76% less

Stock 4,809,500

VBW 1 1,289,425

VBW 2 1,170,000

Stock and VBW 1 are calculated, not benchmarked: lines of loaded files at about 15 tokens each, 80 requests a phase assumed, baseline from an earlier analysis, VBW 1.30.0. Confidence: medium. How we counted, and how →
VBW 2 figures are placeholders until measured; the method will be published beside them.

The rewrite

VBW 2

A program, not the model, decides what counts as done: tests that must fail first, a contract sealed with a hash, and proof on a clean copy of your code.

How it works

The movement