Appearance
Most AI tools still behave like a very fast colleague who is not allowed to touch your computer. You paste text in, you copy text out. The document, the spreadsheet, the folder of half-named PDFs — none of it moves.
Claude Cowork was built to close that gap. When Anthropic shipped the research preview on January 12, 2026, the pitch was blunt: "Claude Code for the rest of your work." The same agent architecture developers drove from a terminal, repackaged for people whose job is contracts, reports and messy folders.
It is also the first mainstream product from a major lab to treat file editing, not conversation, as the core capability — and nine months in, it has been folded into the main Claude app and blamed for one of the ugliest weeks software stocks have had in years.
Table of Contents
- What Is Claude Cowork, Exactly?
- What Cowork Actually Does With Your Files
- Cowork vs Chat vs Claude Code vs a Multi-Model Workspace
- The 60+ Legal Connectors and Why Vertical Data Is the Real Battleground
- Why Software Stocks Fell When Cowork Shipped Plugins
- What Cowork Still Cannot Do
- How to Get the Same File-Working Capability Today
What Is Claude Cowork, Exactly?
Cowork is an autonomous agent that takes a described outcome and works across your files, connected applications and the web until it hands back a finished deliverable. Anthropic's product page reduces it to one instruction: give it a goal, and it works across your files and tools.
You point Claude at a folder, describe the outcome, and it writes a plan, executes step by step, and shows what it is doing: which files it opens, which tools it calls, which choices it makes. You can interrupt or redirect at any point.
The timeline explains the current shape of the product. Research preview in January 2026 for Max subscribers on macOS, Pro four days later, an open-source plugin system on January 30, Windows and enterprise controls in February, general availability on every paid plan on April 9, web and phone in July with cloud sessions that survive a closed laptop. On September 16, Cowork and chat merged into one Claude.
That merge is the part people still get wrong. Anthropic said users were frustrated deciding where a task belonged, and that work started in one surface did not carry into the other. So Cowork stopped being a destination and became a capability Claude reaches for when a request needs files or a multi-step plan. Enterprise plans got it on October 23.
What Cowork Actually Does With Your Files
Inside the folder you grant, Cowork can read, create, edit, rename, sort and delete files. It sorts by content rather than filename, so scanned invoices get grouped by vendor rather than by whatever the scanner called them. It deduplicates, converts formats, pulls text out of images and PDFs, returns Word, Excel, PowerPoint and PDF outputs, runs scheduled tasks, splits big jobs across parallel subagents, and builds live artifacts: dashboards that reconnect to your apps and refresh when reopened.
Anthropic published a figure in July that says more than any feature list: over 90% of Cowork usage was not software development, and business operations plus content creation accounted for roughly half of all usage. People are not using it to prototype. They are using it to get through the document pile.
One caveat about "editing." Cowork edits the file, not the application. It will not sit inside a live Word co-authoring session while three colleagues type, so it is better at what surrounds those sessions: first drafts, redlines against a playbook, extraction across a stack of agreements, the summary nobody wants to write.
Cowork vs Chat vs Claude Code vs a Multi-Model Workspace
They are not interchangeable.
| Dimension | Plain Claude chat | Claude Cowork | Claude Code | Multi-model workspace (Nolvia) |
|---|---|---|---|---|
| Input and output | Prompt in, text out | An outcome plus a folder in, finished files out | A repo and a task in, committed code out | A document in, analysis and structured drafts out |
| File access | None, unless you upload | Read, edit, create, delete in a granted folder | Full project tree | Upload and analyze documents, export the output |
| Best for | Quick questions, thinking out loud | Reports, spreadsheets, decks, file cleanup, recurring ops | Software engineering, multi-step refactors | Document analysis, cross-model comparison, review loops |
| Model choice | Claude only | Claude only | Claude only | 40+ models, one subscription |
| Main limit | Cannot touch your files | Paid plan required; local access needs the desktop app open | Requires a developer's mental model | No autonomous desktop agent |
Chat is a conversation, Cowork is a colleague with a folder, Claude Code is that colleague with a terminal. A multi-model workspace like Nolvia sits on a different axis: which brain answers, not whether the brain can hold a pen. For the coding side, Claude Code with ChatGPT covers how the two fit into one workflow.
The 60+ Legal Connectors and Why Vertical Data Is the Real Battleground
On October 28, 2026, Anthropic shipped a gallery of ready-to-use artifacts for legal work: six it built (cite check, jurisdiction survey matrix, matter board, playbook deviation board, regulatory horizon scan, tabular review), four from Goodwin, Harvey, Foley & Lardner and PwC, and nine access-to-justice tools. The number that matters is buried mid-announcement: since Claude for Legal launched in May 2026, Anthropic has added more than 60 legal connectors, including Bloomberg Law, Filevine, Ivo, Legora and Patlytics.
Why does a connector count matter more than a feature list? Because an agent that can edit files but cannot reach the system of record is a smart intern with no badge. The model is close to commodity; the plumbing is not. Reaching Westlaw, iManage or a firm's own precedent library is where an agent stops being a demo and becomes infrastructure. Legal came first for a reason: Anthropic said legal professionals had become the most engaged Cowork users of any knowledge-work function.
There is a second lesson in the artifacts. Robert Mahari, who leads Claude for Legal, said legal work is not always suited to a pure chat interface. A tracker that sits beside the conversation, that a team can adapt and share with ethical walls and a source link on every output, is a different model from ask-a-question-get-a-paragraph. That pattern — agent, vertical connectors, shareable artifact — is what Anthropic is repeating across industries. If contract review is your actual job, the model comparison for legal contract analysis goes deeper on clause extraction.
Why Software Stocks Fell When Cowork Shipped Plugins
The market shock came in February 2026, not at launch. On February 3, Reuters reported that a selloff across software, data and legal-analytics companies had deepened, with investors pointing at plug-ins Anthropic shipped for Cowork the previous Friday — January 30, the day 11 open-source plugins went up on GitHub. Thomson Reuters, owner of the Westlaw database, fell nearly 18% that session. CNN reported that an ETF tracking the software industry dropped 5.69% that Tuesday, its worst day since April. Wall Street named the episode the SaaSpocalypse.
The panic was not about the plug-ins, which were markdown and JSON files with no compiled code. It was about direction of travel. Most enterprise software is sold per seat, and an agent that does part of several people's jobs is a customer that needs fewer seats. Schroders framed it to Reuters as the loss of a "visibility premium": investors had paid up for predictable subscription revenue, and better agents made long-range forecasts harder to defend.
That is the real meaning of AI moving from chatting with you to doing the work for you. It is not a product story, it is a pricing-model story: when a tool can be described as "a workflow someone bills for," an agent that executes that workflow threatens the invoice, not the feature set. The market has not settled the argument — the iShares Expanded Tech-Software ETF climbed roughly 41% from its April 13 low through October 12, while sitting only about 5% higher for the year.
What Cowork Still Cannot Do
Token economics are brutal. A Cowork task can consume 50 to 100 times the tokens of a single chat message, according to TechTarget's analysis. Agentic work is not free, it is priced differently — hence Enterprise's Efficient mode, model entitlements per role, spend limits and pooled group budgets.
Local file access has a leash. Since October 6, new tasks on Pro and Max run in Anthropic's cloud. Reaching local folders still requires the desktop app to be open. Shut the lid and the local half stops.
Tasks belong to one person. A running task cannot be shared, so teams pass around the output, not the session.
Approval modes are a real decision. Manual, automatic and skip-all exist; auto is the default on Pro, Max and Team, manual on Enterprise. Permanent deletion always requires explicit permission.
Prompt injection is unsolved. Anthropic states that Cowork carries unique risks because of its agentic nature and internet access, and that despite safeguards "the chances of an attack are still non-zero." Its guidance: a dedicated working folder, sensitive files kept out of connected folders, and caution about unfamiliar plugins.
It is Claude-only. If a task suits another engine better — a long-context extraction pass, cheap high-volume classification, a writing voice another model handles more naturally — Cowork will not help. That is a product boundary, not a bug, and it is the gap Nolvia fills.
How to Get the Same File-Working Capability Today
The valuable half of Cowork is not the agentic loop. It is the decision to stop treating AI as a text box. Once the deliverable is a file rather than a paragraph, most of that capability is already reachable without an Enterprise rollout or a $100-plus Claude tier.
The workflow that works on Nolvia, right now:
- Convert first. Drop the PDF, scanned contract or spreadsheet in and ask for structured extraction — parties, dates, obligations, clause numbers. Get text you can actually diff.
- Run the same prompt across models. This is the step a single-model subscription cannot do. Send the identical extraction prompt to Claude, GPT and Gemini, then compare. Engines fail differently on tables, scanned pages and long contracts, and the disagreements point straight at what a human must check.
- Draft in one model, review in another. A model that wrote a summary is a poor judge of it. Hand the draft to a second engine with an explicit review brief.
- Assemble and export. The output is a file, which is the whole point.
Nolvia gives you GPT, Claude, Gemini, Grok and 40+ models in one subscription, with no API keys to manage and no tab juggling. For document-heavy work that means one place where a contract gets read by three engines and the differences get flagged. Pricing is flat and points-based: $15/mo for 45K points, $30/mo for 100K points, $60/mo for 200K points. A 40-page contract read by three models costs three times the tokens of one read by one, so size the tier against actual volume.
What you give up by skipping Cowork: the agent will not rename 4,000 files in your downloads folder while you sleep. What you keep: full model choice, cross-model verification, and no dependency on one vendor's roadmap. For operations, legal-adjacent, consulting and content workflows, that trade is usually the right way around. If you are still mapping the category, a comparison of aggregator platforms and a guide to using multiple models together cover the setup decisions.
Try Every Major AI Model in One Workspace
Nolvia gives you GPT, Claude, Gemini, Grok and 40+ models in a single subscription — no API keys, no juggling five tabs.
Start Free on Nolvia