Tokenless v0.3.2

Patch release to sync the public README with the tested GitHub install path.

Changes:

• Quick start now uses tokenless repair-hooks --user, which works for fresh installs and fixes stale hook paths after upgrading or switching from a local npm link install.
• Verified GitHub install path with npm install -g github:MaxForAI/Tokenless, tokenless --help, and tokenless doctor.

Note: this GitHub release does not publish to npm.

Tokenless v0.3.1

Patch release for the public launch README.

Changes:

• Fixed the top benchmark badge anchor after renaming the section to Benchmarks & Evidence.
• Added localized README files for Chinese, Japanese, French, and Spanish.
• Added first-screen Quick start commands.
• Reworked the benchmark section into Benchmarks & Evidence with real Claude Code API-body results plus related compression and brevity research.
• Restored the faucet logo and added star, license, and acknowledgements sections.

Note: this GitHub release does not publish to npm.

Tokenless v0.3.0

Tokenless v0.3.0 is the first launch-ready release focused on public Claude Code token reduction workflows.

Highlights:

• Public profiles: chat, coding, and off.
• chat shortens everyday Claude responses while staying readable.
• coding uses dense coding-oriented output for faster, lower-token implementation sessions.
• off is now a true hard-off mode that disables both style injection and compression hooks.
• Added picker-friendly Claude Code slash commands: /tokenless-style-chat, /tokenless-style-coding, /tokenless-style-off.
• Added persistent profile config through ~/.tokenless/style.json.
• Polished README, logo, installation docs, benchmark framing, and public release copy.

Benchmarks documented in the repo:

• Vibe coding CRM benchmark: 47.3% fewer request tokens, 44.4% fewer response tokens, 39.3% fewer requests with coding vs off.
• Clean chat benchmark: 80.0% fewer response tokens with chat vs off.

Note: this GitHub release does not publish to npm.

v0.2.1

Patch release

This release aligns the GitHub release tag with npm publishing readiness.

Changed

• Removed the npm private package flag so tokenless can be published.
• Added an npm files whitelist so local API body captures, benchmark scratch directories, and private launcher scripts stay out of published packages.

Verification

• npm publish --dry-run passed for tokenless@0.2.1.
• Package size: 113.4 kB.
• Unpacked size: 536.4 kB.
• Total files: 76.

v0.2.0

Highlights

• Added tokenless launch, a Lean Claude Code launcher that disables high-overhead Task/Plan tools by default while keeping normal read, edit, write, and bash tools available.
• Added TOKENLESS_ALLOW_TASK_TOOLS=1 tokenless launch to restore Claude Code's native task list and plan-mode UI for a session.
• Added true OFF mode via TOKENLESS_MODE=off|false|0|disabled for clean ON/OFF benchmarks.
• Expanded large read packets for JS/TS, Python, Vue, and Svelte files.
• Added large existing Write overwrite protection to steer agents toward bounded edits.
• Added compact edit/write packets for large successful Edit, MultiEdit, and low-risk Write results.

Benchmarks

Measured from raw Claude Code API request bodies:

• Lean launcher Task/Plan trimming: 1,524,894 -> 1,087,753 request tokens, 28.7% reduction.
• Default Lean + Tokenless ON/OFF multifile dashboard: 628,261 -> 512,521 request tokens, 18.4% reduction.
• Large CSS visual edit: about 54-60% request-token reduction.
• 10k-line React/TSX edit: 917,137 -> 545,456 request tokens, 40.5% reduction.

See docs/benchmarking.md for protocol, directories, true-OFF checks, and caveats.

Verification

Passed before release:

• npm run eval:cli-smoke
• npm run eval:read
• npm run eval:edit
• npm run eval:complex
• npm pack --dry-run