Cowork vs OpenClaw: managed agent or self-hosted one

Nobody switches between these two on features. They occupy opposite ends of an operational spectrum: one is a product you subscribe to, the other is a program you operate. Pick the wrong end and you will either pay for control you did not want or inherit a maintenance job you did not budget for.

What is OpenClaw, and what was it called before?

OpenClaw is a free and open-source autonomous agent that executes tasks via large language models and uses messaging platforms as its main interface, released under the MIT licence. It runs as a self-hosted bot connected to services such as Signal, Telegram, Discord and WhatsApp, and works with several model providers rather than one.

The naming history matters when you search for documentation, because most of what is written about this project is filed under a dead name. Per its Wikipedia entry: Warelay (24 November 2025), CLAWDIS (3 December 2025), Clawdbot (2 January 2026), Moltbot (27 January 2026), and OpenClaw since 30 January 2026. Four renames in ten weeks is itself a fact about the project's maturity.

• AI agents
• Self-hosted vs hosted AI tools

How do the two differ operationally?

• Cowork: hosted by the vendor. Anthropic announced web and mobile access on 7 July 2026, with remote sessions running on Anthropic's servers so work continues while your devices are offline. Access comes with a Claude subscription rather than a separate purchase.
• OpenClaw: hosted by you. You provide the machine, the uptime, the updates and the model API keys, and you choose the provider. Nothing runs when your box is off.
• Cowork: the interface is Anthropic's app on desktop, web and mobile. OpenClaw: the interface is a messaging app you already use.
• Cowork: one vendor's models. OpenClaw: whichever provider you configure, which is the main reason people pick it.

What does self-hosting actually cost you here?

Beyond hardware and tokens, it costs a security posture. OpenClaw's documented risk profile includes broad permissions across email, calendars and messaging, susceptibility to prompt-injection attacks, third-party skills that have not been adequately vetted, and misconfigured instances exposed to the internet.

Those are not reasons to avoid it. They are the job description. An agent with shell access and your messaging accounts is a production system, and if nobody on your side owns patching and configuration review, the managed product is the cheaper option even at a subscription price.

• AI tool lock-in: what to check first
• OpenClaw vs Claude Code: scope and hosting

Which one fits which situation?

• You want an agent working on your files and calendar without running infrastructure: Cowork.
• You want provider choice, no per-seat cost, and you are comfortable operating a service: OpenClaw.
• You have a data-residency rule that forbids sending documents to a third party: OpenClaw, configured against a provider your policy already permits.
• You want to hand off a task from a laptop and check it from a phone: Cowork, which is what the web and mobile release was built for.
• You are evaluating for a team with no on-call engineer: the managed product, every time.

What we deliberately do not claim

You will find popularity numbers for OpenClaw quoted everywhere. We do not publish them, because repository star counts move faster than this page can be revised and a stale count is worse than no count. Licence, hosting model and rename history are stable facts; popularity is not.

Neither vendor is affiliated with this site, and product names appear here only to identify the products compared.