On January 30, 2026, Anthropic released 11 plugins for Claude Cowork on GitHub.
Five days later, the market had wiped $285 billion from software companies.
Thomson Reuters plunged 18%, its worst day in history. LegalZoom fell 20%. And then S&P Global, FactSet, SAP, Oracle, Adobe, Salesforce… all bleeding. Traders gave it a name: the “SaaSpocalypse”.
And what does Cowork have to cause such a fall?
It's not another code co-pilot (that's why it's scarier).
Other tools that seemed to play in the same league, such as Cursor, GitHub Copilot and Windsurf, compete for the same thing: helping programmers write code. It’s a huge market, yes, but it’s a market.
This is where Cowork plays a different game.
It is the first AI agent from a large laboratory, Claude, designed for any professional, not just developers.
In practice. Cowork can touch Excel files and insert formulas that actually work, generate full PowerPoint presentations, write decently formatted Word documents, and merge PDFs without breaking them or organize entire folders while you do something else.
Plus, if you combine it with the Claude extension in Chrome (which has been around for a while), it can navigate websites, fill out forms, run multi-step flows in the browser, etc.
And all this is running on your local machine. Your data doesn’t leave your computer because it runs in an isolated virtual machine (you decide which folders it can access).
Underneath runs Claude Opus 4.6, which has 1 million context tokens. Basically, it can process long documents, complex databases, and projects with many files without losing its thread.
Fun fact: according to Simon Willison, the Anthropic team built Cowork using Claude Code in about two weeks.
The plugins that scared Wall Street
Okay, what exactly do those plugins that triggered the panic do?
- Legal reviews contracts and classifies each clause by risk level. Green if it’s okay, yellow if it requires attention, red if it’s problematic. He triages NDAs, runs compliance flows… things that big firms pay junior analysts to do for months.
- Finance generates journal entries, does reconciliations, builds financial statements. And it connects directly to Snowflake, Databricks, BigQuery.
- Sales researches prospects (LinkedIn, news, public reports), prepares meetings, and manages the pipeline by connecting to your CRM.
- And there is one called Plugin Create that lets you build custom plugins from scratch. It basically democratizes the creation of specialized agents.
Can you see why the market panicked?
JP Morgan came out and said it was “an illogical leap” – no one is going to throw away their entire SaaS infrastructure for an Anthropic plugin. Gartner put it more precisely: plugins are disruptors at the task level, but they don’t replace enterprise applications that manage critical operations.
And they are right, of course. But the fear is not because of what it does today. It’s because of the speed at which this is evolving.
What you can really do (and what you can't do)
Alex Finn documented an interesting workflow. Every morning, he writes “Let’s start our day,” and Cowork interviews him about his priorities. Then he decides what he can do alone and launches 3 sub-agents in parallel: one researching the X algorithm (crawling ~100 sites), one analyzing energy actions, and one writing the YouTube strategy.
All in 2 minutes. His conclusion: “Completed 80% of my work for the day in 3 minutes”.
Mike Murphy tried something more mundane but just as useful: renaming 162 podcast transcript files with a naming convention that the Mac Finder doesn’t support. Cowork wrote the script in Python and ran it in 25 seconds. That would have taken half an hour to an hour manually.
Now the limitations, because it has them and they are important:
- Zero memory between sessions. You close the conversation and start everything from scratch. If you worked on something complex yesterday, today it’s time to recontextualize.
- The app must remain open. If you close it, the session dies. There is no way to pause and resume.
- Consumes tokens like a beast. Pro users ($20/month) report reaching the limit quickly with complex tasks. For heavy use you will need the Max plan ($100-200/month).
- Prompt injection vulnerability. Two days after the release, PromptArmor demonstrated that a malicious document could trick Cowork into exfiltrating confidential files. Anthropic is partially patched, but they admit that the risk is not zero.
One of the best articles I’ve found on using Claude Cowork is by Alex Finn, who explains how to make it your daily “operating system.” I’m already putting it into practice, and it’s really amazing.
Compared to what you already know
GitHub Copilot has 42% market share, 20 million users, is in the top 90% of the Fortune 100. It costs $10/month. But it’s exclusively for code.
Cursor has 1 million daily active users, generates over $500M in annual revenue, and offers the best contextual understanding of codebases in the market. $20/month. Also, code only.
Devin promises to work as an autonomous engineer, but independent tests show a 15-30% success rate on complex tasks. Still immature.
The key difference: everyone competes for developers.
Cowork goes for the marketer analyzing campaign data. The lawyer reviewing contracts. The financier reconciling accounts. The product manager synthesizing user research…
The market for knowledge workers is orders of magnitude larger than the market for developers. That’s Anthropic’s bet.
The enterprise issue (which is not minor)
Anthropic has SOC 2 Type II, ISO 27001, ISO 42001, HIPAA configuration, SSO, API compliance certifications. 80-85% of its revenue (run rate of $9B by the end of 2025) comes from enterprise. More than 300,000 enterprises use Claude.
But, crucially, Cowork is specifically not ready for regulated loads.
Cowork activities are not recorded in audited logs. The data is locally on the user’s computer, outside of IT control. There are no role-based access controls. From a compliance standpoint, this is an unbearable headache.
Is it worth it?
For repetitive tasks, file organization, document generation and personal productivity, it already provides real value today.
For regulated work, team collaboration and critical flows, it needs to mature.
What’s interesting is not what it does today, but what it points to: the moment when AI agents ceased to be programmers’ assistants and became what it points to: the moment when AI agents stopped being programmers’ assistants to become any professional’s co-workers.
There is no turning back.
How about you? Are you already thinking about what tasks you would delegate if you had an agent that actually operates on your files?
Leave me your comments.
Have a good week!
