Tokenless
starsforks
How to install Tokenless?
maxforai/tokenless
$git clone https://github.com/MaxForAI/TokenlessOne command to cut token usage by up to 50%+
README.md
| 1 | <p align="center"> |
| 2 | <img src="assets/tokenless-logo.png" alt="Tokenless faucet logo" width="360" /> |
| 3 | </p> |
| 4 | |
| 5 | <h1 align="center">Tokenless</h1> |
| 6 | |
| 7 | <p align="center"> |
| 8 | <strong>One command to cut token usage by up to 50%+.</strong> |
| 9 | </p> |
| 10 | |
| 11 | <p align="center"> |
| 12 | <a href="#benchmarks--evidence"><img alt="vibe coding request reduction" src="https://img.shields.io/badge/vibe%20coding-47.3%25%20less%20request%20tokens-2dd4bf?style=for-the-badge"></a> |
| 13 | <a href="#output-profiles"><img alt="chat response reduction" src="https://img.shields.io/badge/chat-80.0%25%20less%20response%20tokens-4ade80?style=for-the-badge"></a> |
| 14 | <a href="LICENSE"><img alt="license MIT" src="https://img.shields.io/badge/license-MIT-f59e0b?style=for-the-badge"></a> |
| 15 | </p> |
| 16 | |
| 17 | <p align="center"> |
| 18 | <a href="README.md">English</a> · |
| 19 | <a href="README.zh-CN.md">简体中文</a> · |
| 20 | <a href="README.ja.md">日本語</a> · |
| 21 | <a href="README.fr.md">Français</a> · |
| 22 | <a href="README.es.md">Español</a> |
| 23 | </p> |
| 24 | |
| 25 | <p align="center"> |
| 26 | <a href="#before--after">Before / After</a> · |
| 27 | <a href="#quick-start">Quick start</a> · |
| 28 | <a href="#installation">Install</a> · |
| 29 | <a href="#output-profiles">Profiles</a> · |
| 30 | <a href="#benchmarks--evidence">Benchmarks & Evidence</a> · |
| 31 | <a href="#roadmap">Roadmap</a> · |
| 32 | <a href="docs/benchmarking.md">Full benchmark guide</a> |
| 33 | </p> |
| 34 | |
| 35 | --- |
| 36 | |
| 37 | ## Quick start |
| 38 | |
| 39 | ```bash |
| 40 | npm install -g github:MaxForAI/Tokenless |
| 41 | tokenless repair-hooks --user |
| 42 | tokenless install-commands --user |
| 43 | tokenless launch |
| 44 | ``` |
| 45 | |
| 46 | Then use Claude Code normally. Switch profiles anytime: |
| 47 | |
| 48 | ```bash |
| 49 | tokenless style chat # readable, shorter replies |
| 50 | tokenless style coding # dense coding output |
| 51 | tokenless style off # full hard-off |
| 52 | ``` |
| 53 | |
| 54 | Claude Code gets expensive when every log, file read, diff, and long reply keeps getting carried into the next request. |
| 55 | |
| 56 | Tokenless fixes that. |
| 57 | |
| 58 | It keeps the raw evidence on your machine, sends Claude a compact version, and lets you expand the original output only when you need it. |
| 59 | |
| 60 | ## Before / After |
| 61 | |
| 62 | | Normal Claude Code | Claude Code with Tokenless | |
| 63 | | --- | --- | |
| 64 | | Reads a large file or log into future context repeatedly. | Stores the raw output locally and sends a compact packet. | |
| 65 | | Verbose final replies become part of the next request history. | `chat` and `coding` profiles keep replies short. | |
| 66 | | Agent trajectory can grow through repeated exploration and task-plan history. | Launcher trims Task/Plan tools by default; packets reduce large read context. | |
| 67 | |
| 68 | Example large-read replacement: |
| 69 | |
| 70 | | Raw context | Tokenless context | |
| 71 | | --- | --- | |
| 72 | | Full file/log output is carried through API requests. | `TOKENLESS-READ-PACKET/0.1` with artifact id, imports, symbols, snippets, nearby files, and exact expansion commands. | |
| 73 | |
| 74 | ## Why Tokenless |
| 75 | |
| 76 | Claude Code sessions can become expensive because tool outputs, file reads, task-plan history, and verbose assistant replies are repeatedly carried through future API requests. Tokenless targets three sources of growth: |
| 77 | |
| 78 | - Large tool output: test logs, build logs, search results, tree output, diffs, large reads, and large successful edit/write results. |
| 79 | - Agent trajectory overhead: repeated request context, high-overhead Task/Plan tools, and large raw file payloads. |
| 80 | - Response verbosity: optional `chat` and `coding` profiles reduce assistant output tokens. |
| 81 | |
| 82 | ## Benchmarks & Evidence |
| 83 | |
| 84 | Tokenless has two evidence layers: real Claude Code API-body measurements, and external research showing why shorter, denser context can reduce cost without automatically reducing quality. |
| 85 | |
| 86 | ### Real Claude Code benchmark runs |
| 87 | |
| 88 | These are API-body measurements from actual Claude Code sessions. The main metric is estimated request-body or response-body tokens from raw API logs, not local hook-side savings estimates. |
| 89 | |
| 90 | | Scenario | Baseline | Tokenless | Reduction | |
| 91 | | --- | ---: | ---: | ---: | |
| 92 | | 5-turn CRM vibe coding, `off` vs `coding` | 4,697,867 request tokens | 2,476,391 | 47.3% | |
| 93 | | 6-turn natural conversation, `off` vs `chat` | 7,223 response tokens | 1,442 | 80.0% | |
| 94 | | Large CSS visual edit | 1,017,642 request tokens | 403,995-473,354 | ~54-60% | |
| 95 | | 10k-line React/TSX edit | 917,137 request tokens | 545,456 | 40.5% | |
| 96 | | Multifile React dashboard | 628,261 request tokens | 512,521 | 18.4% | |
| 97 | | Task/Plan tools enabled vs default launcher | 1,524,894 request tokens | 1,087,753 | 28.7% | |
| 98 | |
| 99 | The strongest current product benchmark is the 5-turn CRM vibe-coding run: a non-specialist user gave vague iterative product-polish prompts. The public `coding` profile reduced request tokens by 47.3%, response tokens by 44.4%, and request count by 39.3% versus clean `off`. |
| 100 | |
| 101 | The clean natural-conversation run isolates `chat`: no file tools or packet reducers were involved, and response tokens dropped by 80.0%. |
| 102 | |
| 103 | Detailed methodology and raw run notes are in [docs/benchmarking.md](docs/benchmarking.md) and [docs/style-benchmark.md](docs/style-benchmark.md). |
| 104 | |
| 105 | ### Research backing |
| 106 | |
| 107 | The research does not prove Tokenless automatically helps every session. It supports the benchmark pre |
People who liked this also viewed
6 picksType
- hyper-waterfallA human-governed AI coding workflow that distills ephemeral session context into persistent project memory—making work traceable, reviewable, and resumable.WorkflowsJul 2026
- PromaProma brings a seamless general-purpose Agent experience to your workflow. Built for 100× professionals and the proactive Agent era, it supports Claude and Pi runtimes, Feishu group chats, and any LLM provider.WorkflowsJul 2026
- lottiOpen-source private logbook with a local agentic layer. Long-living AI agents read what you record and propose what to do next. Hardware permitting, the models runs locally too. Matrix + Vodozemac for end-to-end encrypted sync between your own devices.WorkflowsJul 2026
- ophelTurn AI conversations into organized, reusable workflows — across major AI platforms. | 把 AI 对话转化为可组织、可复用的工作流，适用于主流 AI 平台WorkflowsJul 2026
- humanize-pptAST-based outline director for human-centered AI presentation workflows.WorkflowsJul 2026
- agent-sessionsLocal-first macOS app to browse, search, analyze, and resume supported AI coding-agent session history across Codex, Claude Code, OpenCode, Cursor Agent, Hermes, OpenClaw, Copilot CLI, and more.WorkflowsJul 2026
