Claude Cowork vs OpenClaw: what changed when Cowork became part of Claude

Claude Cowork is no longer a product you choose: on 16 September 2026 Anthropic merged it into Claude, so the real comparison is now the Claude app against OpenClaw. Pick Claude if you want scheduled agent work that runs on Anthropic's servers with no machine of your own to keep awake, billed as one subscription. Pick OpenClaw if you need to choose your own models, keep the agent on hardware you control, or drive it from Telegram, Signal or WhatsApp.

The rest of this article is the detail behind that answer, re-checked against both vendors' own pages on 20 September 2026.

#What happened to Claude Cowork?

Anthropic's announcement, published on 16 September 2026, says Cowork and chat are now one Claude: quick questions and substantial multi-step work happen in the same conversation instead of in two places. The rollout is staged. Anthropic's blog says it is "rolling out on Pro and Max plans over the next few weeks, with more plans to follow", and the help centre adds that it starts "with Pro and Max plans on web, desktop, and mobile" and that Enterprise organisations get at least 30 days' notice before their accounts move.

Nothing is thrown away in the move. The help centre states that "your existing Cowork tasks, projects, connectors, skills, artifacts, and files carry over", and the blog lists chats, projects, artifacts, connectors and skills as coming with you.

The same announcement introduced Claude Docs and Claude Slides in beta on paid plans, and moved Claude Design inside conversations, also in beta on paid plans. Anthropic's own wording for what you can do with the output is that you can edit, present or download it as PowerPoint or PDF. VentureBeat, reporting the announcement on 16 September 2026, went further and said Claude Docs can be exported to "Microsoft Word or Google Docs, with additional export destinations planned"; that detail is press reporting, not something Anthropic's own pages state, so treat it as unconfirmed until it appears in the help centre. VentureBeat also dates the original Cowork research preview to 12 January 2026, which is again reporting rather than a vendor page.

One naming trap before going further: Microsoft Copilot Cowork is a different product from a different vendor. Everything below is about Anthropic's Claude and the Cowork capabilities now folded into it.

#What is OpenClaw?

OpenClaw is a self-hosted gateway that connects chat apps you already use, among them Discord, Google Chat, iMessage, Matrix, Microsoft Teams, Signal, Slack, Telegram, WhatsApp and Zalo, to AI coding agents. Its documentation is blunt about the commercial model: "No paid tier, no telemetry by default beyond a version check you can turn off, no lab owns it", under an MIT licence. It runs on macOS, Linux, Windows and WSL2, with iOS and Android available as paired nodes. The current release listed on the project's release page is v2026.9.5.

Because you host it, the cost shape is different: the software is free and the bill arrives from whichever model provider you point it at, or from nobody if you run a local model. Our OpenClaw platform hub tracks the setup guides, and the free getting started walkthrough is the shortest path to a running agent.

#How do Claude and OpenClaw compare?

Question	Claude, with Cowork built in	OpenClaw
Where the work runs	Anthropic's servers, plus your own machine when a task needs local files	A gateway you install on your own macOS, Linux or Windows machine
Price	Free at $0; Pro at $17 a month with the annual discount ($200 billed up front) or $20 billed monthly; Max from $100 a month	MIT licensed with no paid tier; you pay your model provider
Scheduled work	Scheduled tasks on all paid plans, running remotely even when your computer is asleep	Built-in scheduler run with openclaw automations, wakes the agent on the host you keep running
Models	Anthropic's models	OpenAI, Anthropic, Google Gemini and Vertex, xAI, OpenRouter, MiniMax, NVIDIA and others, plus local runtimes
Provider failover	Not applicable, since there is one provider	Fallback chains set in agents.defaults.model.fallbacks
Screen control	Computer use, in beta on Pro and Max, on macOS and Windows desktop	Computer use over the gateway's own desktop or a paired node, macOS supported, Windows and Linux experimental
Browser	Built-in browser, tried before screen control	A dedicated agent-only Chrome, Brave, Edge or Chromium profile
How you reach it	Claude on web, desktop and mobile	The chat app you already use, via channel plugins
Parallel work	Sub-agent coordination: Claude splits complex work into smaller tasks and runs parallel workstreams	Sub-agents, background agent runs spawned from an existing run, on a model you choose

#Where does the work actually run?

This is the question that changed most, and an older version of this article got it wrong. Anthropic's scheduling page now states plainly that "scheduled tasks run remotely, so they run on their cadence even when your computer is asleep or the Claude Desktop app is closed", and that scheduled tasks are available on all paid plans. So a recurring Claude task that only touches connected services does not need a machine of yours to be awake.

The exception is local files, and it carries two conditions rather than one. The help centre page on using Cowork across devices says "a cloud session can read and write files in folders you've connected on your computer only while the desktop app is open on that computer and the session was started on desktop". Starting a session on web or mobile therefore does not reach those folders, even with the desktop app running. Read the two pages together and the rule is: the schedule lives in the cloud, and anything reaching into your own folders needs a session you began on desktop, with the desktop app open.

OpenClaw inverts the arrangement. Its scheduler is described as "OpenClaw's built-in scheduler", which "persists jobs, wakes the agent at the right time, and can deliver output to a chat channel, a webhook, or nowhere", managed with openclaw automations (openclaw cron remains an alias). Nothing about that is remote: it runs on the host you chose, so a job fires when that host is on. People usually solve this by putting the gateway on a machine that stays up, which is what the Mac mini setup guide is for.

#Can you choose the model?

On Claude, you use Anthropic's models. That is not a defect so much as the product's shape, and it removes a class of configuration from your life.

OpenClaw's documented provider list is long: OpenAI, Anthropic, an OpenAI ChatGPT and Codex OAuth path, OpenCode, Google Gemini by API key, Google Vertex and the Gemini CLI runtime, Z.AI (GLM), Vercel AI Gateway, OpenRouter, Kilo Gateway, MiniMax, NVIDIA and xAI, with custom providers configurable on top. Local inference is supported through llama.cpp, llmman, LM Studio, Ollama, vLLM and SGLang.

The part worth knowing is failover. OpenClaw builds a model candidate chain from the current selection and a fallback policy, configured through agents.defaults.model.fallbacks, and an agent can override only its fallback chain while still inheriting the shared primary. The docs name what makes it advance to the next candidate: authentication failures, rate limits and cooldown exhaustion, overloaded or provider-busy errors, timeout-shaped failures, billing disables, and model_not_found including eligible HTTP 404 responses. It deliberately does not advance on context overflow errors, which stay inside compaction and retry logic. An explicit model choice you make yourself stays strict and skips the fallback behaviour.

#What can each one do on your screen?

Claude's computer use is in beta for Pro and Max plans, in the Claude Desktop application on macOS and Windows. Team and Enterprise plans do not have access at the time of writing, and Anthropic's download page notes that computer use "is not yet available on Linux, and is coming soon". It is not the first thing Claude reaches for: the help centre describes an order of preference, connectors first, then the browser, then interacting with the screen directly. There is a hard requirement worth planning around, stated on that page: "your computer needs to be awake and the Claude Desktop app needs to be open for computer use to work".

OpenClaw has its own answer, and it is no longer the clean split that earlier comparisons drew. Its docs say "computer use lets an agent see and control the Gateway's own desktop or a paired node's desktop through one built-in computer tool", with screenshots kept model-only and never delivered to the chat channel. It needs a vision-capable model, a tool policy exposing computer, and on macOS both Accessibility and Screen Recording permissions. Windows and Linux support is marked experimental and currently controls only the primary display.

For web work, OpenClaw runs a dedicated Chrome, Brave, Edge or Chromium profile that the agent controls: "the agent opens tabs, reads pages, clicks, and types in this isolated lane", and the docs frame it as "a separate, agent-only browser" whose profile never touches your personal one. Agent actions cover clicking, typing, dragging and selecting, plus snapshots, screenshots and PDFs.

#What does OpenClaw give you that Claude does not?

Running sub-agents is not one of the differences: both do it. Anthropic's help centre lists "Sub-agent coordination: Claude breaks complex work into smaller tasks and coordinates parallel workstreams to complete them", and says that on a complex task "Claude may coordinate multiple sub-agents working simultaneously". OpenClaw defines them as "background agent runs spawned from an existing agent run", used to "parallelize research, long tasks, and slow tool work without blocking the main run". What differs is how much of the arrangement is yours to set. Three things stand out, all of them documented rather than inferred.

Choosing what the sub-agents run on. OpenClaw lets you point sub-agents at a different, cheaper model than the parent, through agents.defaults.subagents.model or a per-agent override. Anthropic's pages document the coordination but no setting for which model those sub-agents use, so on Claude the split is the product's call rather than yours. Our multi-agent workflows guide covers the patterns.

Memory you can read. OpenClaw keeps four memory-related files: an optional USER.md for stable preferences and project context, MEMORY.md for durable long-term facts and decisions, dated daily notes under memory/, and an optional DREAMS.md. The docs state the constraint that makes this trustworthy: "the model only remembers what gets saved to disk; there is no hidden state". The memory system guide goes through the files one at a time.

Reaching the agent where you already are. A Claude conversation happens in Claude. An OpenClaw conversation can happen in Discord, iMessage, Signal, Slack, Telegram or WhatsApp, because the gateway is a channel router first and an agent second.

#What is still rough in the merged Claude?

Anthropic lists the known gaps itself, which is worth more than any outside review. At the time of writing the help centre names these: "Add from GitHub" is unsupported, conversation branching is unavailable, incognito chats open in the previous experience without file and code creation, search leaves out older Cowork tasks, and Dispatch is unavailable to new users.

None of those is fatal, and a staged rollout is exactly when you expect to find them. But if your workflow depended on branching a conversation or on Dispatch, check before you move.

#Which one should you choose?

Choose Claude if:

• You want the shortest path from nothing to a working agent, with no host to maintain.
• Your recurring work runs against connected services rather than local folders, so it can run remotely while your machine is off.
• You want documents, slides and design output in the same place as the conversation.
• One predictable subscription is easier for you to justify than metered model spend.

Choose OpenClaw if:

• You want to pick the model, and to fail over to another provider when one goes down.
• You want the agent reachable from the chat app you already live in.
• You want long-term memory you can open in a text editor and edit.
• You want parallel sub-agents, and you are comfortable being the person who installs, updates and secures the thing.

Run both if you want Claude for document-shaped work on a subscription and OpenClaw as the always-on agent on hardware you own. They are not mutually exclusive, and OpenClaw can call Anthropic's models, so choosing it does not mean giving up Claude.

The honest summary is that the gap narrowed from both directions this year. Claude moved its scheduled work into the cloud and grew a computer-use beta; OpenClaw grew its own computer use, an agent-only browser and a provider failover story. What still separates them is not capability so much as who holds the keys: one is a subscription to somebody else's servers, the other is a daemon on a machine you are responsible for. Pick the trade you would rather live with. The Claude platform hub tracks the first, and the OpenClaw security hardening guide is where to start on the second.