/ Каталог / Песочница / lean-ctx
● Сообщество yvgude ⚡ Сразу

lean-ctx

автор yvgude · yvgude/lean-ctx

Cut AI coding token costs by up to 99% — Rust MCP server with 42 tools for cached reads, context compression, and smart file modes.

lean-ctx is a single Rust binary that reduces AI coding costs through two mechanisms: a shell hook that compresses CLI output (60-95% savings) and an MCP server with 42 tools for cached file reads, context optimization, and multi-agent coordination. Supports 24 AI coding tools (Cursor, Claude Code, Copilot, Windsurf, etc.) with zero telemetry.

Зачем использовать

Ключевые функции

Живое демо

Как выглядит на практике

lean-ctx.replay ▶ готово
0/0

Установка

Выберите клиент

~/Library/Application Support/Claude/claude_desktop_config.json  · Windows: %APPDATA%\Claude\claude_desktop_config.json
{
  "mcpServers": {
    "lean-ctx": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/yvgude/lean-ctx"
      ],
      "_inferred": true
    }
  }
}

Откройте Claude Desktop → Settings → Developer → Edit Config. Перезапустите после сохранения.

~/.cursor/mcp.json · .cursor/mcp.json
{
  "mcpServers": {
    "lean-ctx": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/yvgude/lean-ctx"
      ],
      "_inferred": true
    }
  }
}

Cursor использует ту же схему mcpServers, что и Claude Desktop. Конфиг проекта приоритетнее глобального.

VS Code → Cline → MCP Servers → Edit
{
  "mcpServers": {
    "lean-ctx": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/yvgude/lean-ctx"
      ],
      "_inferred": true
    }
  }
}

Щёлкните значок MCP Servers на боковой панели Cline, затем "Edit Configuration".

~/.codeium/windsurf/mcp_config.json
{
  "mcpServers": {
    "lean-ctx": {
      "command": "TODO",
      "args": [
        "See README: https://github.com/yvgude/lean-ctx"
      ],
      "_inferred": true
    }
  }
}

Тот же формат, что и Claude Desktop. Перезапустите Windsurf для применения.

~/.continue/config.json
{
  "mcpServers": [
    {
      "name": "lean-ctx",
      "command": "TODO",
      "args": [
        "See README: https://github.com/yvgude/lean-ctx"
      ]
    }
  ]
}

Continue использует массив объектов серверов, а не map.

~/.config/zed/settings.json
{
  "context_servers": {
    "lean-ctx": {
      "command": {
        "path": "TODO",
        "args": [
          "See README: https://github.com/yvgude/lean-ctx"
        ]
      }
    }
  }
}

Добавьте в context_servers. Zed перезагружается автоматически.

claude mcp add lean-ctx -- TODO 'See README: https://github.com/yvgude/lean-ctx'

Однострочная команда. Проверить: claude mcp list. Удалить: claude mcp remove.

Сценарии использования

Реальные сценарии: lean-ctx

How to cut your AI coding costs by 80%+ with lean-ctx

👤 Developers paying for Claude, Cursor, or Copilot API tokens ⏱ ~10 min beginner

Когда использовать: Your monthly AI coding bill is growing and you want to reduce it without changing your workflow.

Предварительные требования
  • lean-ctx installed — brew install lean-ctx or curl -fsSL https://leanctx.com/install.sh | sh
Поток
  1. One-command setup
    Run: lean-ctx setup✓ Скопировано
    → Shell hooks installed, all detected editors configured automatically
  2. Code normally for a day
    Use your AI coding tool as usual — lean-ctx works transparently in the background.✓ Скопировано
    → No workflow change; compression happens silently
  3. Check savings
    Run: lean-ctx gain✓ Скопировано
    → Dashboard showing token savings per category (file reads, git, shell, etc.)

Итог: Measurable token reduction (typically 60-90%) with zero workflow friction.

Подводные камни
  • Compressed output confuses some edge-case prompts — Use lean-ctx-off to temporarily disable, or use ctx_read with 'full' mode for specific files

Read large codebases efficiently with 8 file modes

👤 AI coding tool users working on large repos ⏱ ~5 min beginner

Когда использовать: Your AI assistant wastes tokens reading entire files when it only needs the structure or signature.

Поток
  1. Use map mode for overview
    lean-ctx read src/main.rs -m map✓ Скопировано
    → File structure at ~10% of full token cost
  2. Use signatures mode for API
    lean-ctx read src/lib.rs -m signatures✓ Скопировано
    → Just function/type signatures, no bodies
  3. Use diff mode for changes
    lean-ctx read src/main.rs -m diff✓ Скопировано
    → Only lines changed since last read — minimal tokens

Итог: The AI gets exactly the context it needs at a fraction of the token cost.

Подводные камни
  • Aggressive mode strips too much for complex refactoring — Use 'task' mode instead — it preserves context relevant to the current task
Сочетать с: filesystem

Share context between AI agents without re-reading everything

👤 Power users running multiple AI agents on the same project ⏱ ~15 min intermediate

Когда использовать: You're using Claude Code for architecture and Cursor for implementation, and they keep re-reading the same files.

Поток
  1. Build context in agent A
    In Claude Code: analyze the architecture of this project using ctx_overview.✓ Скопировано
    → Context built and cached by lean-ctx
  2. Share to agent B
    In Cursor: use ctx_share to pull the architecture context from the last session.✓ Скопировано
    → Cursor gets the architecture map without re-reading all files
  3. Track costs across agents
    lean-ctx gain --all-sessions✓ Скопировано
    → Combined savings across both agents

Итог: Multiple agents share a cached context layer, avoiding redundant file reads.

Подводные камни
  • Stale cache after file changes — lean-ctx tracks file modification times — cache invalidates automatically on change

Комбинации

Сочетайте с другими MCP — эффект x10

lean-ctx + filesystem

Use lean-ctx for token-efficient reads and filesystem MCP for writes — best of both worlds

Use ctx_read to analyze the codebase structure efficiently, then use filesystem MCP to apply the refactoring changes.✓ Скопировано
lean-ctx + github

lean-ctx compresses local context, GitHub MCP handles remote operations — minimizing total tokens

Use ctx_overview to understand the project, then use GitHub MCP to create a well-informed PR description.✓ Скопировано

Инструменты

Что предоставляет этот MCP

ИнструментВходные данныеКогда вызыватьСтоимость
ctx_read file: str, mode: full|map|signatures|diff|aggressive|entropy|task|lines Read a file with controlled verbosity 0 — local cached read
ctx_multi_read files: str[], mode?: str Read multiple files in one call 0
ctx_tree path: str Get project structure 0
ctx_search query: str, path?: str Find patterns in code 0
ctx_smart_read file: str, intent: str Read with task-aware filtering 0
ctx_shell command: str Run shell commands with output compression 0
ctx_overview path: str Get a high-level project overview 0
ctx_session - Manage context across conversation turns 0
ctx_cost - Track how much you've saved 0

Стоимость и лимиты

Во что обходится

Квота API
No external API — everything runs locally.
Токенов на вызов
That's the point: ctx_read in map mode uses ~10% of the tokens a full read would. Shell compression saves 60-95%.
Деньги
Free and open-source (MIT). Saves $30-100+/month on AI coding API costs for active developers.
Совет
Use 'map' mode as default for exploration, 'full' only when you need every line. The savings compound fast.

Безопасность

Права, секреты, радиус поражения

Хранение учётных данных: No credentials needed. Purely local tool.
Исходящий трафик: Zero. No telemetry, no analytics, no network requests. Everything stays on your machine.

Устранение неполадок

Частые ошибки и исправления

Shell commands produce garbled output

Run lean-ctx setup to update hooks to the latest version. If persists, run lean-ctx-off to disable temporarily.

Проверить: lean-ctx doctor
MCP server not detected by editor

Run lean-ctx setup again — it auto-configures all detected editors. For manual setup: claude mcp add lean-ctx lean-ctx

Проверить: lean-ctx doctor
Binary not found after install

Shell aliases auto-fallback safely. Re-run the install script or brew reinstall lean-ctx.

Проверить: which lean-ctx
ctx_read returns stale content

lean-ctx tracks file mtime for cache invalidation. If you edited outside the normal flow, touch the file to reset the mtime.

Проверить: lean-ctx cache --status

Альтернативы

lean-ctx в сравнении

АльтернативаКогда использоватьКомпромисс
Rust Token KillerYou want a similar concept with different tradeoffsFewer tools (~50 patterns vs 42 tools), 3 editor support vs 24, has default-on telemetry with PII
Manual .cursorrules / CLAUDE.md context managementYou prefer manual context curationNo automatic compression; no caching; more work

Ещё

Ресурсы

📖 Читать официальный README на GitHub

🐙 Открытые задачи

🔍 Все 400+ MCP-серверов и Skills