Moltworker
by Cloudflare
Run a self-hosted personal AI agent on Cloudflare's edge instead of a Mac mini
Moltworker is Cloudflare's open-source middleware for running OpenClaw, the self-hosted personal AI assistant, on Cloudflare Workers instead of dedicated local hardware. It combines a router Worker, a Sandbox container, R2 storage and Zero Trust Access so an always-on personal agent costs roughly $35 a month rather than a machine in a closet.
Moltworker is an open-source middleware project published by Cloudflare on 29 January 2026 that runs OpenClaw — the self-hosted personal AI assistant previously known as Clawdbot and then Moltbot — on the Cloudflare developer platform rather than on a dedicated machine such as the Mac mini its community typically used. The design pairs an entrypoint Worker acting as API router, proxy and administration UI with a Sandbox container that executes the standard OpenClaw gateway runtime, so the agent logic is unmodified while the hosting substrate changes entirely. Persistent state — configuration, session data and conversation history — lives in R2 object storage mounted into the container as a filesystem. Model calls route through Cloudflare AI Gateway, which supplies provider-agnostic routing, unified billing and cost visibility. Browser automation runs on Cloudflare Browser Rendering via a custom Chrome DevTools Protocol proxy exposing Puppeteer APIs, letting the agent drive headless Chromium. The API and admin UI sit behind Cloudflare Zero Trust Access with policy-based login methods and JWT validation. Chat integrations cover Telegram, Discord and Slack. Running it requires a Workers Paid plan for Sandbox container access plus either an Anthropic API key or AI Gateway credentials; AI Gateway is free and R2 and Browser Rendering carry usable free tiers. The repository estimates roughly $34.50 per month for continuous operation, falling substantially with idle-sleep configuration, and notes cold starts of one to two minutes. Cloudflare is explicit that this is a proof of concept, licensed Apache-2.0, not a supported product, and may break without notice.
A developer already running or curious about OpenClaw who wants an always-on personal agent without buying, powering and maintaining dedicated hardware at home.
An always-available personal AI agent with browser automation and chat integrations for roughly $35 a month, with no hardware to own or patch.
At a Glance
- Category
- Agent Development
- Pricing
- Free, Usage-based, Subscription
- Target Market
- Enterprise Developers, Platform Engineers, CTOs, AI Engineers
- Deployment
- Open-source, Cloud-first, Edge-first
- Founded
- 2026
- Headquarters
- San Francisco, United States
- Team Size
- 500+
- Customers
- Not disclosed; the GitHub repository lists 9.9k stars and 1.7k forks
Key Features
- ✓Worker router and admin UI
An entrypoint Worker proxies API traffic and serves an administration interface for managing the agent instance.
- ✓Sandbox container runtime
The standard OpenClaw gateway runs inside a Cloudflare Sandbox, giving isolated execution of untrusted agent code via the Sandbox SDK.
- ✓R2-backed persistence
Object storage mounts as a container filesystem so configuration, session data and conversation history survive restarts.
- ✓AI Gateway model routing
Routes requests to Anthropic and other providers with unified billing, caching and cost visibility across model calls.
- ✓Browser Rendering automation
A custom Chrome DevTools Protocol proxy exposes Puppeteer APIs so the agent can drive headless Chromium for real page tasks.
- ✓Zero Trust Access protection
Access policies and JWT validation guard both the API and the admin UI, so the agent is not publicly exposed.
- ✓Multi-channel chat integrations
Telegram, Discord and Slack connectors let the same agent be reached from the messaging tools people already use.
Capabilities
Use Cases
- •Hardware-free personal agent
Replace a dedicated Mac mini running OpenClaw with an edge deployment that stays available without local power or maintenance.
- •Agent-driven browser tasks
Drive headless Chromium through Browser Rendering to fill forms, gather pages or complete flows that need a real browser.
- •Messaging-first automation
Trigger personal workflows from Telegram, Discord or Slack, with conversation history persisted in R2 across sessions.
- •Reference architecture for agent hosting
Study a working example of sandboxed agent execution behind Zero Trust as a pattern for building supported internal systems.
- •Multi-provider model cost tracking
Use AI Gateway in front of the agent to compare providers and see unified spend across every model call.
Ideal For
Best For
- ✓Running an always-on personal AI assistant without buying or maintaining a dedicated Mac mini or home server
- ✓Developers who want a concrete reference implementation of Workers, Sandboxes, R2, AI Gateway and Zero Trust Access working together
- ✓Prototyping agent architectures on serverless infrastructure before committing to a supported platform
- ✓Personal automation reachable from Telegram, Discord or Slack with browser control for tasks that need a real page
- ✓Learning how to put an untrusted agent runtime behind Zero Trust authentication and sandbox isolation
Not Ideal For
- ✗Production or business-critical workloads — Cloudflare labels it a proof of concept rather than a product, and the repository warns it may break without notice
- ✗Privacy-motivated self-hosters, who lose the original appeal outright: local and home integrations are gone, data now lives in the cloud, and Cloudflare can in principle read everything passing through it
- ✗Anyone connecting an agent to systems they genuinely care about, given the unresolved prompt-injection exposure that commenters repeatedly raised
- ✗Teams needing a support contract or SLA, since there is none for an experimental Apache-2.0 repository
Integrations
Deployment
Market & Ratings
Not disclosed; the GitHub repository lists 9.9k stars and 1.7k forks
Market Analysis
Pros
- ✓Eliminates the dedicated hardware requirement while keeping an always-on agent, at a documented and fairly modest monthly cost
- ✓Layered security done properly for a demo: Zero Trust Access with JWT validation guards the admin UI, and the runtime executes inside an isolated Sandbox
- ✓A genuinely useful reference implementation of six Cloudflare primitives working together, valuable even if you never run the agent
- ✓Well received on Hacker News at 246 points, with the improved Workers Node.js compatibility drawing specific praise
- ✓Open-sourced under Apache-2.0 with setup documentation, so the whole design can be inspected and forked
Cons
- ✗Explicitly a proof of concept and not a Cloudflare product — the repository warns it may break without notice and carries no support or SLA
- ✗It inverts the self-hosting premise: HN commenters noted local and home integrations are gone, data must live in the cloud, and Cloudflare can read everything passing through
- ✗Prompt injection remains unaddressed, with the top-voted concern being that connecting these agents to anything you care about is reckless until that is solved
- ✗Cold starts of one to two minutes, and WebSocket proxying has known limitations in local development
- ✗Commenters criticised the absence of a realistic all-in cost estimate covering every Cloudflare product involved at typical daily usage
- ✗The launch drew accusations of astroturfing and heavy-handed marketing, with several readers reacting to the volume of promotion around it
Pricing
Open source (Apache-2.0)
$0
- ✓Full source on GitHub
- ✓Setup documentation
- ✓Telegram, Discord and Slack integrations
- ✓No support or SLA
Cloudflare Workers Paid (required)
From $5/mo
- ✓Sandbox container access
- ✓10M Worker requests/month included
- ✓Free AI Gateway
- ✓R2 and Browser Rendering free allocations
Moltworker itself is free and Apache-2.0, but running it is not. A Cloudflare Workers Paid plan is mandatory because Sandbox containers are gated behind it, starting at $5/month minimum. On top of that you pay container compute — Cloudflare bills containers at $0.0000025 per GiB-second of memory and $0.000020 per vCPU-second — and the repository's own estimate for continuous 24/7 operation is roughly $34.50 per month, broken down as about $26 memory, $2 CPU, $1.50 disk plus the $5 plan fee. Idle-sleep configuration reduces this substantially. Model inference is billed separately by Anthropic against your own API key. AI Gateway is free and R2 and Browser Rendering include usable free tiers.
Security & Compliance
Connect
Sources
This page was written from 5 sources, 4 on domains other than github.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Kitesurf
An agent-first browser that runs in V8 isolates on Cloudflare Workers
Natural
Payments infrastructure that lets AI agents hold, send and collect money
AgentPrizm
Governed, persistent memory and reusable skills for AI agents — over REST API and MCP
Bespoke Labs
Reinforcement-learning environments that train reliable, production-ready AI agents