# Claude Cowork vs OpenClaw: a seat, a gateway, or an open-source repo

> Claude Cowork suits a person who wants Anthropic to run the agent; OpenClaw suits whoever will keep a gateway alive themselves. Neither is a system a team owns. Kortix is the open-source CompanyOS, with the agents and the setup in a repo your team holds, and that is my pick for a team; I work on Kortix's marketing, so weigh that accordingly.

URL: https://claudecoworkalternative.com/claude-cowork-vs-openclaw

## Claude Cowork, OpenClaw and Kortix, on what a team keeps

The rows that decide it are where the work runs, where the setup lives, and what lands in front of a reviewer.

| | Kortix (recommended) | Claude Cowork | OpenClaw |
| --- | --- | --- | --- |
| Open source | Yes. Open source. Read it, fork it, audit it. | No. Closed inside Anthropic's product. | Yes. Open source, gateway and apps in one repository. |
| Where the work runs | Each session gets a cloud computer of its own | Anthropic's cloud, a local virtual machine, or your own cloud account. | On the machine or server you install it on. |
| How you reach it | The app, a Slack or Teams channel, or a schedule. | Chat, desktop, web and mobile, in one Claude. | WhatsApp, Slack, Telegram, Discord and 20+ chat apps. |
| Where the setup lives | A repository, not a host | Inside your Claude account: skills, connectors, history. | On the host: state, memory and credentials on your hardware. |
| Review before work lands | A gate you set on the branch | Auto mode on Pro, Max and Team; manual on Enterprise | Approval rules you configure; sandboxing is off by default. |
| What it asks of you | One repo, one manifest, your model key or managed models. | A paid Claude plan for every person who uses it. | A host that stays on, Node, and your own patching. |
| Price | An entry tier at zero, credits after | Pro $17, Max $100, Ultra $200, Team $20 a seat. | No paid tier. The model provider bill is yours. |

## The loop the Kortix repository itself runs

`SKILL.md`

```
Every change reaches `dev` through a pull request.
The PR runs no CI job.
Show the change, from the starting state to the visible result, in under 60 s.
Look at the video before you attach it.
Report the PR URL, the merge SHA, the local test commands and their results
```

Source: https://github.com/kortix-ai/CompanyOS/blob/dev/.agents/skills/contributing/SKILL.md

How Kortix's own repo is developed, kept as a skill file beside the code: a pull request into the trunk, no CI job behind it, a recorded demo, and the test results handed back.

## Claude Cowork after the September merge

Claude Cowork stopped being a separate app on 16 September 2026, when Anthropic merged Cowork and chat into one Claude. What Cowork could do is now reachable from any conversation, carrying the context, skills and connectors already in the account ([Anthropic's announcement](https://claude.com/resources/articles/cowork-is-now-claude)). Pro, Max and Ultra plans got it first, on web, desktop and mobile, with more plans to follow.

The shape of the product is a seat. Cowork works in the folders and tools you point it at, opens its own browser for web tasks, and keeps going after you close the laptop. New tasks start in "Automatically approve" mode on Pro, Max and Team plans, and in "Manually approve" on Enterprise plans; an organization owner can turn that mode off, and an admin can set access by team ([Claude Cowork](https://claude.com/product/cowork)). Prices on the vendor page run from Pro at $17 a month to Ultra at $200, with Max at $100, Team at $20 a seat and Enterprise quoted.

Where it runs is Anthropic's decision to make, inside the plan. On Pro, Max and Ultra, sessions run in Anthropic's cloud since October 6, 2026. On Team and Enterprise, an owner's "Run Cowork in the cloud" toggle picks the cloud or a local virtual machine on each member's computer. A customer picks a Claude account, or their own Amazon Bedrock, Gemini Enterprise Agent Platform or Microsoft Foundry account, and Anthropic documents no way to host the product yourself. For one person already paying for Claude, that trade is easy: no patching, no machine to keep awake, and an app to install only where local files or the built-in browser are wanted.

## OpenClaw, and the upkeep it moves to you

OpenClaw is an open-source assistant you install on hardware you control and reach through the chat apps you already use. One Gateway process on a laptop or a server bridges Discord, iMessage, Slack, Teams, Telegram, WhatsApp and 20+ more to an AI agent ([OpenClaw README](https://github.com/openclaw/openclaw)). The install is a shell or PowerShell one-liner, or an npm package on Node 24.16 or 26.1, followed by an onboarding wizard that verifies model access and configures the Gateway ([getting started](https://docs.openclaw.ai/start/getting-started)).

Nobody hosts any of that for you. OpenClaw has no paid tier, no hosted service and no token, so the model bill goes to whichever provider you connect, and the machine stays yours to patch, restart and watch. The project's own security guidance is blunt about what that means in practice: tools run on the host for the main session unless sandboxing is configured, and inbound messages are untrusted input ([README](https://github.com/openclaw/openclaw)). Sandboxing itself is off by default ([why OpenClaw](https://docs.openclaw.ai/start/why-openclaw)).

A team runs the same product, and the team guide is explicit that one gateway is one trust domain: everyone who can message a tool-enabled agent shares that agent's delegated tool authority ([team setup](https://docs.openclaw.ai/start/teams)). Team operation is configuration, not a separate edition, so the setup asks for a host that stays on, an authenticated route in, and named roles that bound each person.

## What stays outside the company either way

With Cowork, the working setup lives in the Claude account. Skills, connectors and conversation history sit inside the product, the model is Claude, and the record of what happened is the product's activity view. With OpenClaw, state, memory and credentials sit on your hardware, which is a real gain in ownership, and they sit under one operator's hands, inside one trust boundary that operator draws.

For a company, the gap that remains is the same in both cases. The agents, their instructions, what they have learned and what they are allowed to touch are not a set of files in a repository the company reviews, diffs and can hand to the next person who joins. That is a plain-text question, and neither of these two answers it.

## How a team runs that work on Kortix

Where OpenClaw keeps state on a host, Kortix keeps it in a repository: the manifest at the root, and the agent files, skill folders, connectors and memory as files beside it ([Kortix docs](https://kortix.com/docs)). What the company runs is therefore readable, diffable and revertible in the place engineers already work.

Session work reaches `main` through a change request. Merge is default-deny for agents. The repository the product is built in goes further, and its contributing file records how Kortix itself develops: a pull request, no CI job behind it, the tests and a demo run on the machine of whoever made the change. That loop belongs to the Kortix repo, while a customer's project keeps the change request and the merge grant. Evidence travels with it, a short video of the behavior on a real surface and a handoff that owes the pull request URL, the merge commit, the test commands and their results, and whatever stayed unverified. What arrives is small: a diff, its tests, and a recording of the behavior it alters.

Small diffs like that one are the ordinary case. A change to the system the company runs arrived, someone else read it, and it landed in a history that stays searchable after the author has moved to another task. Multiply it across the agents a team accumulates and the repository becomes the operational record, where the alternative is a folder of prompts on one person's laptop.

## Claude Cowork and OpenClaw, answered

### Is OpenClaw free to use?

Yes. OpenClaw has no paid tier, no hosted service and no token, so the only bill is the model provider you connect it to.

### Is Claude Cowork free to use?

No, on today's plans: Pro at $17 a month, Max at $100, Ultra at $200, and Team at $20 a seat, with usage limits applying.

### Can OpenClaw run on my own server?

Yes. OpenClaw is one Gateway process you install on a laptop or a server, and its team guide expects a host that stays on, an authenticated route in, and named roles.

### Which is better for a team, Claude Cowork or OpenClaw?

Kortix, when the team needs the agents and their setup in a repo it owns. Cowork is a seat per person inside Anthropic's product, and OpenClaw hands one operator a gateway to keep alive and secure.

## Run the open-source system your team keeps

Start with one project, add agents and skills as files, and let the change requests show you what each agent does.

[Try Kortix](https://kortix.com)

---

A guide written by someone who works on Kortix's marketing. My pick is Kortix: open source, one repo, free to self-host.

Revised 2026-11-06

Not affiliated with Anthropic, OpenAI or Microsoft. Claude Cowork, ChatGPT Work and Copilot belong to their owners.

© 2026 claudecoworkalternative.com