Claude Cowork
Files, spreadsheets, documents without a terminal; works in a folder on your machine, the model runs in Anthropic's cloud.
Verified: Sep 13last change: Aug 28next price check: Dec 12
Facts with sources
- It reaches local folders through the Claude Desktop app, with permissions scoped to specific folders. Source (in Russian)
- The same technology is sold by Microsoft inside Microsoft 365 as Copilot Cowork. Source (in Russian)
- Unavailable from Russia, along with Claude. Source (in Russian)
Strengths
- Reads a whole folder at once: Richard Yu fed it 400 pages of emails, pages and call transcripts and it found 3 self-contradictions.
- In Sadie St Lawrence's test it caught a hidden $18,000 stipend line and the CFO's 150-participant requirement that ChatGPT Work missed.
- Access is limited to folders you grant, once or always; before risky steps like deleting files it pauses and shows what it is about to do.
- Scheduled tasks run daily, weekly or monthly unattended; Richard Yu's morning automation ran without him touching it.
Weaknesses
- Memory is per project: Richard Yu kept a rules copy in each one and after 6 weeks had 5 diverging versions, none clearly current.
- Claude chat and Cowork learn nothing from each other, so the same correction gets explained twice on the same account.
- Anything that needs the browser is slow: it looks at the screen and clicks step by step, the way a person would.
- Handles simple column-based spreadsheets well but struggles with merged cells and unusual layouts.
Example from practice
An executive memo from three conflicting files
Sadie St Lawrence gave Claude Cowork and ChatGPT Work the same packet, an executive brief, leadership meeting notes and a budget spreadsheet, with no web access, and asked for a decision memo, a five-slide deck and an action tracker. The files deliberately disagreed on headcount (120 vs 150), launch date and budget: the sheet showed $133,100 under a $150,000 ceiling, but the notes required a $150 per-person stipend, making the real total $151,100. Both then received the same CFO revision. Cowork won on her weighted score and applied the correction; ChatGPT Work missed the requirement. Elapsed time is not stated in the video.
source: Sadie St Lawrence, 'ChatGPT Work vs Claude Cowork: I Gave Both the Same Executive Job'
What reviewers say
Reviewer opinions on their own tasks and hardware, not measurements. Links go to the videos; transcripts live in docs/raw-data.
- I'm never opening Claude Cowork again · Richard Yu AI, Aug 15After 6 weeks of client work the reviewer moved to Claude Code for shared memory, but says Cowork is the right choice for contained, safer file work.
- Claude Cowork Explained: What It Actually Is (For Total Beginners) · IFF, Aug 16The reviewer's point is that it does the work instead of explaining it: folder sorting, a sourced research summary and a 340-row spreadsheet cleanup were demonstrated.
- ChatGPT Work vs Claude Cowork: I Gave Both the Same Executive Job · Sadie St Lawrence, Aug 18The reviewer gave Cowork the win for staying correct rather than for polish, while stressing the human still owns the decision.
- Claude for Word: Turn a draft into a finished document · Claude, Aug 26Vendor demo: first summarise reviewer comments and name conflicts, turn on track changes, check every number against the source, list explicitly what not to touch.
- Delegate and schedule tasks in Claude Cowork · Claude, Jun 21Vendor demo: name sources and output format explicitly, corrections can be added mid-task; in June scheduled tasks ran only with the desktop app open.
- The future of work with @Claude · Claude, Jul 2The vendor on Claude Tag in Slack: give the agent a standing objective per channel, public channels let the team copy expert users; "65% of PRs written by the agent" is an internal figure.
What it can do for you
- Pull together a project summary from a folder of reports and emails
- Merge three spreadsheets into one and flag the discrepancies
- Draft a presentation from a set of documents
Models in this agent
Score is from our manager-task benchmark; models without a score are not measured yet.
What changed
- Claude Cowork gained a built-in browser: it opens sites, fills forms and collects data on its own →
- Claude for Word: the agent edits a document with track changes on and checks numbers against the source →
- Claude Cowork: delegation and scheduled tasks; in June they ran only while the desktop app was open →
How to brief an agent like this
The four-line brief is the Foundation module: prompt structure, LLM agents and MCP, an AI policy for the team. It starts with the free open module.