What is Claude Cowork — and how to export your agent tasks
If you've opened Claude lately you've probably noticed a new toggle next to "Chat": Cowork. Flip it, hand Claude a goal — "pull NVIDIA's latest annual report and summarize it," "generate these four files" — and instead of a reply you get an agent run: Claude spins up a cloud sandbox, runs commands, searches the web, writes code, and delivers finished files to an Outputs panel. It's Claude Code in a friendly coat, available to everyone.
Which raises the question every powerful Claude feature eventually does: how do you keep it? A Cowork task can be a real piece of work — a research trail, a generated report, a small pile of files. And there is, of course, no "Export" button.
What Cowork actually is (and why it's different)
A Cowork task isn't a chat and isn't a project. It lives at a URL like claude.ai/cowork/cse_<id>, and under the hood it's a Claude Code session — an append-only log of the agent's steps: your prompts, Claude's replies, every tool it ran (Bash, file edits, web search) and what each returned, plus the files it handed back. That's a different shape from a normal conversation, which is exactly why the usual "save this chat" paths don't reach it.
- A chat (
/chat/<id>) is a plain conversation. - A project (
/cowork/project/<id>) groups chats with shared instructions and knowledge files. - A Cowork task (
/cowork/cse_<id>) is an agent run that does things and produces files.
How to export a Cowork task
Open the task, click the Export button AI Chat Archive puts on the page (on a Cowork task it slides aside so it never covers Claude's Progress / Outputs / Context panel), pick a format — Markdown, HTML, PDF, JSON or TXT — and you're done. Step by step: how to export a Claude Cowork task.
What lands in your downloads folder:
- The transcript — your prompts and Claude's replies, plus a readable trace of the tools the task ran and their results (the agentic work, not a wall of raw logs). Internal scaffolding is filtered out. This part is free, in every format.
- The output files — on a paid plan, every file the task delivered to its Outputs panel (PDFs, spreadsheets, Word docs, images, text) is downloaded with its real bytes and bundled into the ZIP under
outputs/, right next to the transcript.
Long tasks export in full — the extension pages through the entire event history, so a fifty-step run doesn't get truncated to its tail. Non-Latin content and filenames (Chinese included) survive end to end, ZIP and all.
Where does any of this go? Nowhere.
Same promise as the rest of the extension: the transcript is rendered and the files are fetched in your browser, using the claude.ai session you're already logged into, and written straight to your downloads folder. No server we run ever sees your task — no account, no upload, no telemetry. The privacy policy names every network destination the extension ever talks to.
The honest summary
Cowork is where Claude stops answering and starts doing — and the output of "doing" is worth keeping. Want just the record of a task? The free transcript export covers it in every format. Made files you actually need? The paid tier bundles them in alongside — a one-time license, not a subscription. Either way your agent runs become something you hold on your own machine, instead of something living in a panel that might look different next week.