Function Compute：使用 OpenClaw Sandbox

前置條件

• 已完成 SDK 接入（E2B API Key、api_url、domain）

• 已按 OpenClaw 模板 構建出狀態為 ready 的模板，並記錄模板名稱（下文以 TEMPLATE_NAME 表示）與構建樣本中的 OPTS

• 已準備模型 API Key，建立沙箱時通過 envs 注入 — 鏡像內不預置

  • 國際站：百鍊控制台 擷取 API Key（sk- 開頭）

部署 Gateway

啟動 OpenClaw Gateway，在瀏覽器中與 Agent 對話。以下完整流程：建立 Sandbox → 配置 → 啟動 Gateway → 擷取訪問地址。服務端編程調用見 Agent CLI。

import time

from e2b_code_interpreter import Sandbox

# Gateway 鑒權 Token：自行設定任一字元串，無需從 API 擷取。
# 啟動時傳給 openclaw gateway --token；瀏覽器訪問時通過 URL ?token= 攜帶。
# 官方文檔亦支援 OPENCLAW_APP_TOKEN 環境變數，本質相同。
TOKEN = "my-gateway-token"
PORT = 18789

BAILIAN_API_KEY = "<YOUR-BAILIAN-API-KEY>"
BAILIAN_BASE_URL = "[workspace-id].[region].maas.aliyuncs.com"
BAILIAN_MODEL = "qwen3.8-max"

# 1. 建立 Sandbox
sandbox = Sandbox.create(
    template=TEMPLATE_NAME,
    timeout=3600,
    envs={
        "ANTHROPIC_AUTH_TOKEN": BAILIAN_API_KEY,
        "ANTHROPIC_BASE_URL": BAILIAN_BASE_URL,
        "ANTHROPIC_MODEL": BAILIAN_MODEL,
    },
    **OPTS,
)

# 註冊百鍊 provider
sandbox.commands.run(
    "openclaw onboard --non-interactive --accept-risk --skip-health "
    "--auth-choice custom-api-key "
    f"--custom-api-key {BAILIAN_API_KEY} "
    f"--custom-base-url {BAILIAN_BASE_URL} "
    "--custom-compatibility anthropic "
    f"--custom-model-id {BAILIAN_MODEL} "
    "--custom-provider-id bailian",
    timeout=120,
)

# 2. 設定預設模型
sandbox.commands.run(
    f"openclaw config set agents.defaults.model.primary bailian/{BAILIAN_MODEL}"
)

# 3. 配置 Control UI（雲沙箱公網訪問必需）
origin = f"https://{sandbox.get_host(PORT)}"
sandbox.commands.run(
    f"openclaw config set gateway.controlUi.allowedOrigins '[\"{origin}\"]'"
)

# 4. 啟動 Gateway（後台運行）
sandbox.commands.run(
    f"bash -lc 'openclaw config set gateway.controlUi.allowInsecureAuth true && "
    f"openclaw config set gateway.controlUi.dangerouslyDisableDeviceAuth true && "
    f"openclaw gateway --allow-unconfigured --bind lan --auth token "
    f"--token {TOKEN} --port {PORT}'",
    background=True,
)

# 5. 等待 Gateway 就緒
for _ in range(45):
    probe = sandbox.commands.run(
        f'bash -lc \'ss -ltn | grep -q ":{PORT} " && echo ready || echo waiting\''
    )
    if probe.stdout.strip() == "ready":
        break
    time.sleep(1)

url = f"https://{sandbox.get_host(PORT)}/?token={TOKEN}"
print(f"Gateway: {url}")

Agent CLI

在服務端以編程方式調用 Agent，無需啟動 Gateway。

若已按上文 部署 Gateway 建立了 Sandbox，直接在同一個 sandbox 上執行下方命令即可，無需再次建立。

若僅需 CLI、不啟動 Gateway：

from e2b_code_interpreter import Sandbox

BAILIAN_API_KEY = "<YOUR-BAILIAN-API-KEY>"
BAILIAN_BASE_URL = "[workspace-id].[region].maas.aliyuncs.com"
BAILIAN_MODEL = "qwen3.8-max"

sandbox = Sandbox.create(
    template=TEMPLATE_NAME,
    timeout=3600,
    envs={
        "ANTHROPIC_AUTH_TOKEN": BAILIAN_API_KEY,
        "ANTHROPIC_BASE_URL": BAILIAN_BASE_URL,
        "ANTHROPIC_MODEL": BAILIAN_MODEL,
    },
    **OPTS,
)

# 註冊百鍊 provider
sandbox.commands.run(
    "openclaw onboard --non-interactive --accept-risk --skip-health "
    "--auth-choice custom-api-key "
    f"--custom-api-key {BAILIAN_API_KEY} "
    f"--custom-base-url {BAILIAN_BASE_URL} "
    "--custom-compatibility anthropic "
    f"--custom-model-id {BAILIAN_MODEL} "
    "--custom-provider-id bailian",
    timeout=120,
)

result = sandbox.commands.run(
    f'openclaw agent --local --agent main --model bailian/{BAILIAN_MODEL} '
    '-m "What is 2+2? Reply with just the number."',
    timeout=120,
)
print(result.stdout)

瀏覽器訪問（雲沙箱）

要在瀏覽器中開啟雲沙箱的 Gateway Control UI，須先綁定自訂網域名。預設平台網域名稱（*.sandbox.aliyuncs.com）會觸發下載而非開啟頁面；綁定自訂網域名後可正常訪問。控制台添加網域名稱、DNS 解析、HTTPS 認證及 SDK 配置的完整步驟，見 雲沙箱自訂網域名。

OPTS = {
    "api_key": "<你的 E2B API Key>",
    "api_url": "https://api.cn-beijing.sandbox.aliyuncs.com",
    "domain": "cn-beijing.sandbox.aliyuncs.com",
}

OPTS = {
    "api_key": "<你的 E2B API Key>",
    "api_url": "https://api.<你的自訂網域名>",
    "domain": "<你的自訂網域名>",
}

curl -sI "https://<host>/?token=<Gateway Token>"

安全模式

測試階段可關閉裝置配對（上文已設定 dangerouslyDisableDeviceAuth true）。若啟用安全模式，開啟 URL 後需批准待配對裝置：

import json

for _ in range(30):
    try:
        res = sandbox.commands.run(
            f"openclaw devices list --json --url ws://127.0.0.1:{PORT} --token {TOKEN}"
        )
        data = json.loads(res.stdout)
        if data.get("pending"):
            rid = data["pending"][0]["requestId"]
            sandbox.commands.run(
                f"openclaw devices approve {rid} --token {TOKEN} "
                f"--url ws://127.0.0.1:{PORT}"
            )
            print(f"Device approved: {rid}")
            break
    except Exception:
        pass
    time.sleep(2)

重啟 Gateway

修改模型或配置後，在當前 Sandbox 中執行：

origin = f"https://{sandbox.get_host(PORT)}"
sandbox.commands.run(
    f"openclaw config set gateway.controlUi.allowedOrigins '[\"{origin}\"]'"
)

sandbox.commands.run(
    """bash -lc 'for p in "[o]penclaw gateway" "[o]penclaw-gateway"; do
for pid in $(pgrep -f "$p" || true); do kill "$pid" 2>/dev/null || true; done
done'"""
)
time.sleep(1)

sandbox.commands.run(
    f"openclaw gateway --allow-unconfigured --bind lan --auth token "
    f"--token {TOKEN} --port {PORT}",
    background=True,
)

for _ in range(45):
    probe = sandbox.commands.run(
        f'bash -lc \'ss -ltn | grep -q ":{PORT} " && echo ready || echo waiting\''
    )
    if probe.stdout.strip() == "ready":
        break
    time.sleep(1)

關閉不安全配置

sandbox.commands.run(
    "bash -lc 'openclaw config set gateway.controlUi.allowInsecureAuth false && "
    "openclaw config set gateway.controlUi.dangerouslyDisableDeviceAuth false'"
)

sandbox.commands.run(
    """bash -lc 'for p in "[o]penclaw gateway" "[o]penclaw-gateway"; do
for pid in $(pgrep -f "$p" || true); do kill "$pid" 2>/dev/null || true; done
done'"""
)

sandbox.commands.run(
    f"openclaw gateway --allow-unconfigured --bind lan --auth token "
    f"--token {TOKEN} --port {PORT}",
    background=True,
)

使用 Skill 擴充能力

OpenClaw 通過 Skill（目錄 + SKILL.md）擴充 Agent 能力，Gateway 與 Agent CLI 均可使用。詳見 OpenClaw Skills。

以下樣本假設已按上文建立 sandbox（Gateway 或 Agent CLI 均可）。鏡像不預裝自訂 Skill；可用 openclaw skills list 查看已載入 Skill（含 bundled）。路徑約定見 OpenClaw 模板。

以下以阿里雲百鍊為例（BAILIAN_MODEL 見上文）。

寫入 Managed Skill：

sandbox.files.write(
    "/home/user/.openclaw/skills/summarize-changes/SKILL.md",
    """---
name: summarize-changes
description: Summarizes uncommitted changes and flags risks. Use when reviewing diffs or writing commit messages.
user-invocable: true
---

## Instructions
1. Run `git diff HEAD` and summarize changes in 2–3 bullets.
2. List risks such as missing error handling or hardcoded values.
3. If the diff is empty, say there are no uncommitted changes.
""",
)

# 確認已出現在列表中
print(sandbox.commands.run("openclaw skills list").stdout)

result = sandbox.commands.run(
    f'openclaw agent --local --agent main --model bailian/{BAILIAN_MODEL} '
    '--message "/summarize-changes"',
    timeout=0,
)
print(result.stdout)

寫入 Workspace Skill：

sandbox.files.write(
    "/home/user/.openclaw/workspace/skills/api-conventions/SKILL.md",
    """---
name: api-conventions
description: API design conventions. Use when adding or changing HTTP endpoints or /healthz.
user-invocable: true
---

When writing API endpoints:
- Use RESTful naming
- Return consistent error formats
- Include request validation
""",
)

result = sandbox.commands.run(
    f'openclaw agent --local --agent main --model bailian/{BAILIAN_MODEL} '
    '--message "/api-conventions Add a /healthz endpoint"',
    timeout=0,
)
print(result.stdout)

從本機上傳多檔案 Skill 目錄：

若 Skill 含 scripts/、references/ 等附屬檔案，可以用 files.write_files 一次批量寫入：

from pathlib import Path

def upload_skill_dir(sandbox, local_dir: str, remote_root: str) -> None:
    local = Path(local_dir).resolve()
    entries = []
    for path in local.rglob("*"):
        if path.is_file():
            rel = path.relative_to(local).as_posix()
            entries.append({
                "path": f"{remote_root.rstrip('/')}/{rel}",
                "data": path.read_bytes(),
            })
    sandbox.files.write_files(entries)

# 本機目錄結構樣本：
# ./my-skills/echo-marker/SKILL.md
# ./my-skills/echo-marker/scripts/marker.txt
upload_skill_dir(
    sandbox,
    "./my-skills/echo-marker",
    "/home/user/.openclaw/skills/echo-marker",  # Managed；Workspace 則用 workspace/skills/...
)

print(sandbox.commands.run("openclaw skills list").stdout)

result = sandbox.commands.run(
    f'openclaw agent --local --agent main --model bailian/{BAILIAN_MODEL} '
    '--message "/echo-marker"',
    timeout=0,
)
print(result.stdout)

關於 Template.copy：暫不支援Template.build 時用 .copy("local-dir", "/home/user/.openclaw/skills/...") 預置檔案。自訂 Skill 請用上方運行時 files.write / files.write_files。

從 ClawHub 安裝（可選）：

沙箱預設具備公網出站，可直接搜尋 / 安裝社區 Skill（瀏覽 clawhub.ai）：

# 搜尋
result = sandbox.commands.run("openclaw skills search calendar", timeout=120)
print(result.stdout)

# 安裝
sandbox.commands.run("openclaw skills install ws-calendar", timeout=180)

參數說明

清理

任務完成後釋放資源：

sandbox.kill()

計費說明

Sandbox 按 CPU、記憶體規格與運行時間長度計費；模型 API 呼叫費用由模型服務單獨結算。詳見 計費概述。

常見問題

相關文檔

• 雲沙箱自訂網域名

• E2B OpenClaw Gateway

• OpenClaw Control UI

• OpenClaw 專案

• OpenClaw 模板：鏡像地址、構建 Template、Gateway 參數、環境概覽、Skill 能力約定

• Base URL總覽