Luban

A side-bar coding agent for VS Code and VSCodium. Describe what you want — it edits files in your workspace, runs a script or the build to verify, then tells you what changed. No patched editor, no takeover of the native chat: a normal extension that installs from the Marketplace.

中文简介：装在侧边栏里的编码 agent。说人话 → 它改文件、跑命令、自己验证，然后告诉你改了什么。零源码补丁，可从 VS Code 市场 / Open VSX / VSIX 安装，也能用脚本装成 App 的内置扩展。完整中文说明 →

Highlights

• Plan first, execute after you approve — Plan mode is read-only by construction (write tools are not registered, commands are checked). The plan is saved as a Markdown file you can edit; executing it re-reads that file.
• Review in the editor, not in chat — every write opens a real diff: accept or reject hunk by hunk from CodeLens or the diff title bar.
• Context without ceremony — @file, @folder/, #selection, plus automatic context from the editor around your cursor; a local BM25 index backs search_code.
• Any OpenAI-compatible provider — DeepSeek, Ollama, llama.cpp, vLLM and friends via ~/.luban/providers.json. API keys go to SecretStorage, never to disk.
• Sessions that survive a reload — history on disk, several concurrent tabs, a message queue while a turn is running, opt-in multi-turn memory.
• Built for cheap multi-turn — message layout is aligned to token prefix caching (typically 95%+ cache hits after the second turn); each turn shows input / output / cache / context usage, and a local ledger feeds a token dashboard.
• Extensible — local stdio MCP servers and SKILL.md skills join the same tool loop; Plan mode only sees tools declared read-only.
• Optional tab completion — DeepSeek FIM completion, off by default.

Install

Requirements: VS Code / VSCodium 1.100+, plus at least one provider with an API key.

Quick start

1. Configure a provider — click ⚙ at the bottom of the panel (or run Provider settings), fill in base URL, models and API key, then hit Test connection.
2. Open the panel — Luban icon in the activity bar, or Cmd+Shift+L.
3. Ask for something — e.g. 新建 utils.js，导出 debounce(fn, ms) 和 throttle(fn, ms)，然后用 node 测一下两者行为.

Editing ~/.luban/providers.json by hand works too (run Reload providers.json after):

{
  "defaultModel": "deepseek::deepseek-flash",
  "providers": [
    {
      "id": "deepseek",
      "name": "DeepSeek",
      "baseUrl": "https://api.deepseek.com/v1",
      "apiKeyEnv": "DEEPSEEK_API_KEY",
      "models": [
        { "id": "deepseek-flash", "name": "DeepSeek Flash",
          "maxInputTokens": 65536, "maxOutputTokens": 8192, "toolCalling": true }
      ]
    }
  ]
}

Modes

Documentation

The Marketplace only renders this page, so the details live on GitHub:

Safety

run_command executes in your own shell — there is no OS-level sandbox. Plan mode adds a heuristic read-only check on commands, and file tools may touch paths outside the workspace (luban.fs.allowOutsideWorkspace=false restricts them). MCP servers run as child processes with your user's permissions: only configure servers you trust. See known limitations.

License

Apache-2.0 © 2026 eppen (liuhualiang). See LICENSE and NOTICE.

The Luban name and icon are not licensed under Apache-2.0.