Container Use MCP — 安裝 & 即時演示

為什麼要用

核心特性

One isolated container per agent task — no host pollution
Multiple agents work in parallel on the same repo via git worktrees inside containers
Replay, diff, and discard any environment after the fact
Pre-baked language images (node, python, go) or BYO Dockerfile
Works with Claude Code, Cursor, Goose, Aider through the same MCP layer

即時演示

實際使用效果

container-use-mcp.replay ▶ 就緒

0/0

安裝

選擇你的客戶端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "container-use-mcp": {
      "command": "cu",
      "args": [
        "mcp"
      ]
    }
  }
}

開啟 Claude Desktop → Settings → Developer → Edit Config。儲存後重啟應用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "container-use-mcp": {
      "command": "cu",
      "args": [
        "mcp"
      ]
    }
  }
}

Cursor 使用與 Claude Desktop 相同的 mcpServers 格式。專案級設定優先於全域。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "container-use-mcp": {
      "command": "cu",
      "args": [
        "mcp"
      ]
    }
  }
}

點擊 Cline 側欄中的 MCP Servers 圖示，然後選 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "container-use-mcp": {
      "command": "cu",
      "args": [
        "mcp"
      ]
    }
  }
}

格式與 Claude Desktop 相同。重啟 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "container-use-mcp",
      "command": "cu",
      "args": [
        "mcp"
      ]
    }
  ]
}

Continue 使用伺服器物件陣列，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "container-use-mcp": {
      "command": {
        "path": "cu",
        "args": [
          "mcp"
        ]
      }
    }
  }
}

加入 context_servers。Zed 儲存後熱重載。

claude mcp add container-use-mcp -- cu mcp

一行命令搞定。用 claude mcp list 驗證，claude mcp remove 移除。

使用場景

實戰用法： Container Use

Try three refactoring approaches in parallel, keep the best

👤 Engineers exploring design space ⏱ ~15 min intermediate

何時使用： You have 3 different ideas for a refactor and want to see all of them run/pass tests before picking one.

前置條件

Server/skill installed and authenticated — See repo README

步驟

Spin up three isolated envs

Create three container-use environments named refactor-a/b/c, all on this repo at HEAD.✓ 已複製

→ Three env_ids printed
Apply each approach in its own env

In refactor-a do the strategy-pattern version; in -b inline everything; in -c extract a service object. Run the full test suite in each. Report pass/fail and diff size.✓ 已複製

→ Per-env test summary and diff stats
Pick the winner and merge

Apply the diff from refactor-b to my working copy. Discard the other two environments.✓ 已複製

→ Patch applied locally

結果： Three real experiments evaluated side-by-side without ever breaking your host repo.

注意事項

Forgetting to discard old envs leaks disk — Forgetting to discard old envs leaks disk — cu environment list then cu environment discard <id> periodically.

搭配使用： github · filesystem

組合

與其他 MCP 搭配，撬動十倍槓桿

container-use-mcp + github

Pair with github for complementary capabilities

Use this server together with github to complete a multi-step task.✓ 已複製

container-use-mcp + filesystem

Pair with filesystem for complementary capabilities

Use this server together with filesystem to complete a multi-step task.✓ 已複製

工具

此 MCP 暴露的能力

工具	輸入參數	何時呼叫	成本
environment_create	name: str, image: str	Start a fresh sandbox for the next task	Local CPU/disk
environment_run_cmd	env_id, cmd	Execute build/test commands inside the sandbox	free
environment_file_write	env_id, path, contents	Apply edits without touching the host	free
environment_diff	env_id	Review the agent's work before merging back	free

成本與限制

運行它的成本

API 配額: See provider docs for rate limits
每次呼叫 Token 數: Varies by tool
費用: See repo README for pricing details
提示: Cache tool results and avoid repeated identical calls.

安全

權限、密鑰、影響範圍

憑證儲存： Use environment variables; never commit secrets

資料出站： Tool calls go to the provider's API as documented

故障排查

常見錯誤與修復

docker daemon not running

Container Use requires a local Docker or compatible runtime. Start Docker Desktop or colima.

驗證： Run `docker ps` successfully

out of disk

Old environments aren't auto-pruned. Run cu environment list then cu environment discard <id> for anything older than a day.

驗證： Check `docker system df`

替代方案

Container Use 對比其他方案

替代方案	何時用它替代	權衡
Dagger Cloud	You want hosted, multi-tenant agent envs	Paid; container-use is local-first

Container Use