OpenClaw on Cloudflare Workers

Run OpenClaw (formerly Moltbot, formerly Clawdbot) personal AI assistant in a Cloudflare Sandbox.

Experimental: This is a proof of concept demonstrating that OpenClaw can run in Cloudflare Sandbox. It is not officially supported and may break without notice. Use at your own risk.

Requirements

• Workers Paid plan ($5 USD/month) — required for Cloudflare Sandbox containers
• Anthropic API key — for Claude access, or you can use AI Gateway's Unified Billing

The following Cloudflare features used by this project have free tiers:

• Cloudflare Access (authentication)
• Browser Rendering (for browser navigation)
• AI Gateway (optional, for API routing/analytics)
• R2 Storage (optional, for persistence)

Container Cost Estimate

This project uses a standard-1 Cloudflare Container instance (1/2 vCPU, 4 GiB memory, 8 GB disk). Below are approximate monthly costs assuming the container runs 24/7, based on Cloudflare Containers pricing:

Resource	Provisioned	Monthly Usage	Included Free	Overage	Approx. Cost
Memory	4 GiB	2,920 GiB-hrs	25 GiB-hrs	2,895 GiB-hrs	~$26/mo
CPU (at ~10% utilization)	1/2 vCPU	~2,190 vCPU-min	375 vCPU-min	~1,815 vCPU-min	~$2/mo
Disk	8 GB	5,840 GB-hrs	200 GB-hrs	5,640 GB-hrs	~$1.50/mo
Workers Paid plan					$5/mo
Total					~$34.50/mo

Notes:

• CPU is billed on active usage only, not provisioned capacity. The 10% utilization estimate is a rough baseline for a lightly-used personal assistant; your actual cost will vary with usage.
• Memory and disk are billed on provisioned capacity for the full time the container is running.
• To reduce costs, configure SANDBOX_SLEEP_AFTER (e.g., 10m) so the container sleeps when idle. A container that only runs 4 hours/day would cost roughly ~$5-6/mo in compute on top of the $5 plan fee.
• Network egress, Workers/Durable Objects requests, and logs are additional but typically minimal for personal use.
• See the instance types table for other options (e.g., lite at 256 MiB/$0.50/mo memory or standard-4 at 12 GiB for heavier workloads).

What is OpenClaw?

OpenClaw (formerly Moltbot, formerly Clawdbot) is a personal AI assistant with a gateway architecture that connects to multiple chat platforms. Key features:

• Control UI - Web-based chat interface at the gateway
• Multi-channel support - Telegram, Discord, Slack
• Device pairing - Secure DM authentication requiring explicit approval
• Persistent conversations - Chat history and context across sessions
• Agent runtime - Extensible AI capabilities with workspace and skills

This project packages OpenClaw to run in a Cloudflare Sandbox container, providing a fully managed, always-on deployment without needing to self-host. Optional R2 storage enables persistence across container restarts.

Architecture

Quick Start

Cloudflare Sandboxes are available on the Workers Paid plan.

# Install dependencies
npm install

# Set your API key (direct Anthropic access)
npx wrangler secret put ANTHROPIC_API_KEY

# Or use Cloudflare AI Gateway instead (see "Optional: Cloudflare AI Gateway" below)
# npx wrangler secret put CLOUDFLARE_AI_GATEWAY_API_KEY
# npx wrangler secret put CF_AI_GATEWAY_ACCOUNT_ID
# npx wrangler secret put CF_AI_GATEWAY_GATEWAY_ID

# Generate and set a gateway token (required for remote access)
# Save this token - you'll need it to access the Control UI
export MOLTBOT_GATEWAY_TOKEN=$(openssl rand -hex 32)
echo "Your gateway token: $MOLTBOT_GATEWAY_TOKEN"
echo "$MOLTBOT_GATEWAY_TOKEN" | npx wrangler secret put MOLTBOT_GATEWAY_TOKEN

# Deploy
npm run deploy

After deploying, open the Control UI with your token:

https://your-worker.workers.dev/?token=YOUR_GATEWAY_TOKEN

Replace your-worker with your actual worker subdomain and YOUR_GATEWAY_TOKEN with the token you generated above.

Note: The first request may take 1-2 minutes while the container starts.

Important: You will not be able to use the Control UI until you complete the following steps. You MUST:
1. Set up Cloudflare Access to protect the admin UI
2. Pair your device via the admin UI at /_admin/

You'll also likely want to enable R2 storage so your paired devices and conversation history persist across container restarts (optional but recommended).

Setting Up the Admin UI

To use the admin UI at /_admin/ for device management, you need to:

• Enable Cloudflare Access on your worker
• Set the Access secrets so the worker can validate JWTs

1. Enable Cloudflare Access on workers.dev

The easiest way to protect your worker is using the built-in Cloudflare Access integration for workers.dev:

• Go to the Workers & Pages dashboard
• Select your Worker (e.g., moltbot-sandbox)
• In Settings, under Domains & Routes, in the workers.dev row, click the meatballs menu (...)
• Click Enable Cloudflare Access
• Copy the values shown in the dialog (you'll need the AUD tag later). Note: The "Manage Cloudflare Access" link in the dialog may 404 — ignore it.
• To configure who can access, go to Zero Trust in the Cloudflare dashboard sidebar → Access → Applications, and find your worker's application:

• Add your email address to the allow list

• Or configure other identity providers (Google, GitHub, etc.)
• Copy the Application Audience (AUD) tag from the Access application settings. This will be your CF_ACCESS_AUD in Step 2 below

2. Set Access Secrets

After enabling Cloudflare Access, set the secrets so the worker can validate JWTs:

# Your Cloudflare Access team domain (e.g., "myteam.cloudflareaccess.com")
npx wrangler secret put CF_ACCESS_TEAM_DOMAIN

# The Application Audience (AUD) tag from your Access application that you copied in the step above
npx wrangler secret put CF_ACCESS_AUD

You can find your team domain in the Zero Trust Dashboard under Settings > Custom Pages (it's the subdomain before .cloudflareaccess.com).

3. Redeploy

npm run deploy

Now visit /_admin/ and you'll be prompted to authenticate via Cloudflare Access before accessing the admin UI.

Alternative: Manual Access Application

If you prefer more control, you can manually create an Access application:

• Go to Cloudflare Zero Trust Dashboard
• Navigate to Access > Applications
• Create a new Self-hosted application
• Set the application domain to your Worker URL (e.g., moltbot-sandbox.your-subdomain.workers.dev)
• Add paths to protect: /_admin/*, /api/*, /debug/*
• Configure your desired identity providers (e.g., email OTP, Google, GitHub)
• Copy the Application Audience (AUD) tag and set the secrets as shown above

Local Development

For local development, create a .dev.vars file with:

DEV_MODE=true               # Skip Cloudflare Access auth + bypass device pairing
DEBUG_ROUTES=true           # Enable /debug/* routes (optional)

Authentication

By default, moltbot uses device pairing for authentication. When a new device (browser, CLI, etc.) connects, it must be approved via the admin UI at /_admin/.

Device Pairing

• A device connects to the gateway
• The connection is held pending until approved
• An admin approves the device via /_admin/
• The device is now paired and can connect freely

This is the most secure option as it requires explicit approval for each device.

Gateway Token (Required)

A gateway token is required to access the Control UI when hosted remotely. Pass it as a query parameter:

https://your-worker.workers.dev/?token=YOUR_TOKEN
wss://your-worker.workers.dev/ws?token=YOUR_TOKEN

Note: Even with a valid token, new devices still require approval via the admin UI at /_admin/ (see Device Pairing above).

For local development only, set DEV_MODE=true in .dev.vars to skip Cloudflare Access authentication and enable allowInsecureAuth (bypasses device pairing entirely).

Persistent Storage (R2)

By default, moltbot data (configs, paired devices, conversation history) is lost when the container restarts. To enable persistent storage across sessions, configure R2:

1. Create R2 API Token

• Go to R2 > Overview in the Cloudflare Dashboard
• Click Manage R2 API Tokens
• Create a new token with Object Read & Write permissions
• Select the moltbot-data bucket (created automatically on first deploy)
• Copy the Access Key ID and Secret Access Key

2. Set Secrets

# R2 Access Key ID
npx wrangler secret put R2_ACCESS_KEY_ID

# R2 Secret Access Key
npx wrangler secret put R2_SECRET_ACCESS_KEY

# Your Cloudflare Account ID
npx wrangler secret put CF_ACCOUNT_ID

To find your Account ID: Go to the Cloudflare Dashboard, click the three dots menu next to your account name, and select "Copy Account ID".

How It Works

R2 storage uses a backup/restore approach for simplicity:

On container startup:

• If R2 is mounted and contains backup data, it's restored to the moltbot config directory
• OpenClaw uses its default paths (no special configuration needed)

During operation:

• A cron job runs every 5 minutes to sync the moltbot config to R2
• You can also trigger a manual backup from the admin UI at /_admin/

In the admin UI:

• When R2 is configured, you'll see "Last backup: [timestamp]"
• Click "Backup Now" to trigger an immediate sync

Without R2 credentials, moltbot still works but uses ephemeral storage (data lost on container restart).

Container Lifecycle

By default, the sandbox container stays alive indefinitely (SANDBOX_SLEEP_AFTER=never). This is recommended because cold starts take 1-2 minutes.

To reduce costs for infrequently used deployments, you can configure the container to sleep after a period of inactivity:

npx wrangler secret put SANDBOX_SLEEP_AFTER
# Enter: 10m (or 1h, 30m, etc.)

When the container sleeps, the next request will trigger a cold start. If you have R2 storage configured, your paired devices and data will persist across restarts.

Admin UI

Access the admin UI at /_admin/ to:

• R2 Storage Status - Shows if R2 is configured, last backup time, and a "Backup Now" button
• Restart Gateway - Kill and restart the moltbot gateway process
• Device Pairing - View pending requests, approve devices individually or all at once, view paired devices

The admin UI requires Cloudflare Access authentication (or DEV_MODE=true for local development).

Debug Endpoints

Debug endpoints are available at /debug/* when enabled (requires DEBUG_ROUTES=true and Cloudflare Access):

• GET /debug/processes - List all container processes
• GET /debug/logs?id=<process_id> - Get logs for a specific process
• GET /debug/version - Get container and moltbot version info

Optional: Chat Channels

Telegram

npx wrangler secret put TELEGRAM_BOT_TOKEN
npm run deploy

Discord

npx wrangler secret put DISCORD_BOT_TOKEN
npm run deploy

Slack

npx wrangler secret put SLACK_BOT_TOKEN
npx wrangler secret put SLACK_APP_TOKEN
npm run deploy

Optional: Browser Automation (CDP)

This worker includes a Chrome DevTools Protocol (CDP) shim that enables browser automation capabilities. This allows OpenClaw to control a headless browser for tasks like web scraping, screenshots, and automated testing.

Setup

• Set a shared secret for authentication:

npx wrangler secret put CDP_SECRET
# Enter a secure random string

• Set your worker's public URL:

npx wrangler secret put WORKER_URL
# Enter: https://your-worker.workers.dev

• Redeploy:

npm run deploy

Endpoints

Endpoint	Description
GET /cdp/json/version	Browser version information
GET /cdp/json/list	List available browser targets
GET /cdp/json/new	Create a new browser target
WS /cdp/devtools/browser/{id}	WebSocket connection for CDP commands

All endpoints require authentication via the ?secret=<CDP_SECRET> query parameter.

Built-in Skills

The container includes pre-installed skills in /root/clawd/skills/:

cloudflare-browser

Browser automation via the CDP shim. Requires CDP_SECRET and WORKER_URL to be set (see Browser Automation above).

Scripts:

• screenshot.js - Capture a screenshot of a URL
• video.js - Create a video from multiple URLs
• cdp-client.js - Reusable CDP client library

Usage:

# Screenshot
node /root/clawd/skills/cloudflare-browser/scripts/screenshot.js https://example.com output.png

# Video from multiple URLs
node /root/clawd/skills/cloudflare-browser/scripts/video.js "https://site1.com,https://site2.com" output.mp4 --scroll

See skills/cloudflare-browser/SKILL.md for full documentation.

Optional: Cloudflare AI Gateway

You can route API requests through Cloudflare AI Gateway for caching, rate limiting, analytics, and cost tracking. OpenClaw has native support for Cloudflare AI Gateway as a first-class provider.

AI Gateway acts as a proxy between OpenClaw and your AI provider (e.g., Anthropic). Requests are sent to https://gateway.ai.cloudflare.com/v1/{account_id}/{gateway_id}/anthropic instead of directly to api.anthropic.com, giving you Cloudflare's analytics, caching, and rate limiting. You still need a provider API key (e.g., your Anthropic API key) — the gateway forwards it to the upstream provider.

Setup

• Create an AI Gateway in the AI Gateway section of the Cloudflare Dashboard.
• Set the three required secrets:

# Your AI provider's API key (e.g., your Anthropic API key).
# This is passed through the gateway to the upstream provider.
npx wrangler secret put CLOUDFLARE_AI_GATEWAY_API_KEY

# Your Cloudflare account ID
npx wrangler secret put CF_AI_GATEWAY_ACCOUNT_ID

# Your AI Gateway ID (from the gateway overview page)
npx wrangler secret put CF_AI_GATEWAY_GATEWAY_ID

All three are required. OpenClaw constructs the gateway URL from the account ID and gateway ID, and passes the API key to the upstream provider through the gateway.

• Redeploy:

npm run deploy

When Cloudflare AI Gateway is configured, it takes precedence over direct ANTHROPIC_API_KEY or OPENAI_API_KEY.

Choosing a Model

By default, AI Gateway uses Anthropic's Claude Sonnet 4.5. To use a different model or provider, set CF_AI_GATEWAY_MODEL with the format provider/model-id:

npx wrangler secret put CF_AI_GATEWAY_MODEL
# Enter: workers-ai/@cf/meta/llama-3.3-70b-instruct-fp8-fast

This works with any AI Gateway provider:

Provider	Example CF_AI_GATEWAY_MODEL value	API key is...
Workers AI	workers-ai/@cf/meta/llama-3.3-70b-instruct-fp8-fast	Cloudflare API token
OpenAI	openai/gpt-4o	OpenAI API key
Anthropic	anthropic/claude-sonnet-4-5	Anthropic API key
Groq	groq/llama-3.3-70b	Groq API key

Note: CLOUDFLARE_AI_GATEWAY_API_KEY must match the provider you're using — it's your provider's API key, forwarded through the gateway. You can only use one provider at a time through the gateway. For multiple providers, use direct keys (ANTHROPIC_API_KEY, OPENAI_API_KEY) alongside the gateway config.

Workers AI with Unified Billing

With Unified Billing, you can use Workers AI models without a separate provider API key — Cloudflare bills you directly. Set CLOUDFLARE_AI_GATEWAY_API_KEY to your AI Gateway authentication token (the cf-aig-authorization token).

Legacy AI Gateway Configuration

The previous AI_GATEWAY_API_KEY + AI_GATEWAY_BASE_URL approach is still supported for backward compatibility but is deprecated in favor of the native configuration above.

All Secrets Reference

Secret	Required	Description
CLOUDFLARE_AI_GATEWAY_API_KEY	Yes	Your AI provider's API key, passed through the gateway (e.g., your Anthropic API key). Requires CF_AI_GATEWAY_ACCOUNT_ID and CF_AI_GATEWAY_GATEWAY_ID
CF_AI_GATEWAY_ACCOUNT_ID	Yes	Your Cloudflare account ID (used to construct the gateway URL)
CF_AI_GATEWAY_GATEWAY_ID	Yes	Your AI Gateway ID (used to construct the gateway URL)
CF_AI_GATEWAY_MODEL	No	Override default model: provider/model-id (e.g. workers-ai/@cf/meta/llama-3.3-70b-instruct-fp8-fast). See Choosing a Model
ANTHROPIC_API_KEY	Yes	Direct Anthropic API key (alternative to AI Gateway)
ANTHROPIC_BASE_URL	No	Direct Anthropic API base URL
OPENAI_API_KEY	No	OpenAI API key (alternative provider)
AI_GATEWAY_API_KEY	No	Legacy AI Gateway API key (deprecated, use CLOUDFLARE_AI_GATEWAY_API_KEY instead)
AI_GATEWAY_BASE_URL	No	Legacy AI Gateway endpoint URL (deprecated)
CF_ACCESS_TEAM_DOMAIN	Yes	Cloudflare Access team domain (required for admin UI)
CF_ACCESS_AUD	Yes	Cloudflare Access application audience (required for admin UI)
MOLTBOT_GATEWAY_TOKEN	Yes	Gateway token for authentication (pass via ?token= query param)
DEV_MODE	No	Set to true to skip CF Access auth + device pairing (local dev only)
DEBUG_ROUTES	No	Set to true to enable /debug/* routes
SANDBOX_SLEEP_AFTER	No	Container sleep timeout: never (default) or duration like 10m, 1h
R2_ACCESS_KEY_ID	No	R2 access key for persistent storage
R2_SECRET_ACCESS_KEY	No	R2 secret key for persistent storage
CF_ACCOUNT_ID	No	Cloudflare account ID (required for R2 storage)
TELEGRAM_BOT_TOKEN	No	Telegram bot token
TELEGRAM_DM_POLICY	No	Telegram DM policy: pairing (default) or open
DISCORD_BOT_TOKEN	No	Discord bot token
DISCORD_DM_POLICY	No	Discord DM policy: pairing (default) or open
SLACK_BOT_TOKEN	No	Slack bot token
SLACK_APP_TOKEN	No	Slack app token
CDP_SECRET	No	Shared secret for CDP endpoint authentication (see Browser Automation)
WORKER_URL	No	Public URL of the worker (required for CDP)

Security Considerations

Authentication Layers

OpenClaw in Cloudflare Sandbox uses multiple authentication layers:

• Cloudflare Access - Protects admin routes (/_admin/, /api/*, /debug/*). Only authenticated users can manage devices.

• Gateway Token - Required to access the Control UI. Pass via ?token= query parameter. Keep this secret.

• Device Pairing - Each device (browser, CLI, chat platform DM) must be explicitly approved via the admin UI before it can interact with the assistant. This is the default "pairing" DM policy.

Troubleshooting

npm run dev fails with an Unauthorized error: You need to enable Cloudflare Containers in the Containers dashboard

Gateway fails to start: Check npx wrangler secret list and npx wrangler tail

Config changes not working: Edit the # Build cache bust: comment in Dockerfile and redeploy

Slow first request: Cold starts take 1-2 minutes. Subsequent requests are faster.

R2 not mounting: Check that all three R2 secrets are set (R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY, CF_ACCOUNT_ID). Note: R2 mounting only works in production, not with wrangler dev.

Access denied on admin routes: Ensure CF_ACCESS_TEAM_DOMAIN and CF_ACCESS_AUD are set, and that your Cloudflare Access application is configured correctly.

Devices not appearing in admin UI: Device list commands take 10-15 seconds due to WebSocket connection overhead. Wait and refresh.

WebSocket issues in local development: wrangler dev has known limitations with WebSocket proxying through the sandbox. HTTP requests work but WebSocket connections may fail. Deploy to Cloudflare for full functionality.

Known Issues

Windows: Gateway fails to start with exit code 126 (permission denied)

On Windows, Git may check out shell scripts with CRLF line endings instead of LF. This causes start-openclaw.sh to fail with exit code 126 inside the Linux container. Ensure your repository uses LF line endings — configure Git with git config --global core.autocrlf input or add a .gitattributes file with * text=auto eol=lf. See #64 for details.

Links

• OpenClaw
• OpenClaw Docs
• Cloudflare Sandbox Docs
• Cloudflare Access Docs