Claude Cowork

Claude Cowork is basically the same application as Claude Desktop with another mode on top of it.

Linux Setup

As of March 2026, there is no official Linux support, but the unofficial port from https://github.com/johnzfitch/claude-cowork-linux works surprisingly well. The only downside I noticed is that file locations are very messy, with some paths coming from the macOS convention for instance.
On Gentoo, emerge dev-util/claude-code and app-arch/7zip (app-arch/p7zip won't work), and make sure the nodejs ebuild is built with the npm USE flag. Then installation is easy.

A skill file is basically a condensed set of instructions that the agent should read at the beginning of a session to understand the tasks he's supposed to achieve. One way to produce a correct skill file is to first train the agent in a session, then ask him at the end to write the skill file from the conversation and the steps he learnt.
A skill file is lightweight; be careful though that a session can get quite big and it's better to break large tasks or projects into separate sessions, write smaller skill files and ask for a merge at the end.

On Linux, the Claude in Chrome extension works by opening a websocket to the Anthropic servers (and Claude Cowork will also open a websocket to them), not via a native messaging host. This means the communication actually takes place via a third party (the Anthropic server). It usually works well, but be careful never to answer "No and don't ask again" if Claude either asks you to install Claude in Chrome. It can do that even if the extension is already installed, and after that this blocks Claude into thinking the extension is not working / not connected. The solution is to remove this setting from the configuration file if it happens.
The Google Drive connector UI is very misleading. On Claude Cowork, even once properly installed and authentified, it still requires you to click the + icon in a conversation (session), click on Connectors and toggle the Drive search on. Else it will not work.

Currently it refuses to save or enter passwords on websites.
It cannot upload files from your system via the standard browser dialog mechanism.

It seems very limited so far. Impossible to write on a Google sheet for instance, and even searching for Google sheet files did not work. Note that on the other hand, Claude can use the Google Sheets API very well.

Be careful than a session can get quite big (because the agent may receive a lot of verbose output from API endpoints for instance). Once it gets larger than 30MB, it's dangerous - in one case the conversation was completely stuck with a (wrong) error message about rate limits being reached.
For this reason, better to break large sessions into smaller ones, even though you may always ask Claude to recover from a blocked session by reading the conversation logs.