Claude Is Becoming a Platform

How Cowork, plugins, and MCP apps are turning a chatbot into work infrastructure.

Over the past five weeks, Anthropic has shipped three things in rapid succession: an autonomous desktop agent called Cowork, a set of eleven open-source plugins that give that agent specialized professional skills, and a collection of interactive app integrations that embed tools like Slack, Figma, and Asana directly inside the Claude interface. Each announcement got its own news cycle, its own round of commentary, and its own moment of market panic (the plugins alone contributed to a $285 billion selloff in software stocks). But taken together, they tell a more coherent story than any single launch suggests: Anthropic is building a platform, and the chatbot is becoming the least interesting part of it.

Cowork: The Agent That Works For You

Cowork launched on macOS on January 12 and arrived on Windows on February 10, which brought it to roughly 70 percent of the desktop market. The basic pitch is straightforward: point Claude at a folder, describe what one wants done, and step away. Cowork reads files, creates new ones, coordinates parallel subtasks, and delivers finished output, whether that means a reconciled expense spreadsheet, a reorganized project directory, or a first-draft report assembled from scattered notes.

Anthropic describes it as “Claude Code for the rest of your work,” which is apt if slightly underselling the ambition. Claude Code, the company’s terminal-based coding agent, is by some accounts generating $1 billion in revenue and has become popular enough that Microsoft now expects its engineers to use it alongside GitHub Copilot. Cowork takes the same underlying architecture and wraps it in a graphical interface that does not require anyone to open a terminal.

The Windows release also brought a pricing shift that matters. When Cowork debuted on Mac, it required a Claude Max subscription at $100 to $200 per month, a filter that selected for early adopters and power users. The Windows version is available on Claude Pro at $20 per month, which puts it in the same price range as most SaaS tools it might replace. That is not an accident.

What Cowork actually does well, based on early user reports, is the boring stuff: sorting files, extracting data from documents, building spreadsheets with working formulas, and chaining together multi-step tasks that would otherwise require a person to babysit each transition. What it does less well is anything requiring judgment calls about ambiguous content, or navigating file systems outside its sandboxed folder. (On Windows, it currently restricts access to the user’s personal folder, which has frustrated some users but strikes others as a sensible guardrail.) It remains in “research preview,” with Anthropic explicitly warning that it can take destructive actions if given unclear instructions.

The Plugin Layer: Teaching Claude Your Job

Two and a half weeks after Cowork launched, Anthropic released eleven open-source plugins covering productivity, sales, finance, legal, marketing, customer support, data analysis, enterprise search, project management, software development, and biology research. Each plugin bundles domain-specific skills, slash commands, data connectors, and sub-agents into a package that turns Claude from a generalist into something closer to a role-specific colleague.

The structure is worth understanding because it explains what spooked investors. A sales plugin, for instance, hooks into HubSpot, Clay, and ZoomInfo through Model Context Protocol (MCP) connectors, then encodes a company’s actual sales process so that Claude can handle prospect research, call preparation, and follow-up drafting without being re-instructed each time. A legal plugin does clause-by-clause contract review with risk flagging, NDA triage, and redline generation based on a firm’s own negotiation playbook. The legal plugin in particular drew sharp reactions: Thomson Reuters shares fell roughly 16 percent in the days following the announcement, and Morgan Stanley analysts wrote that Anthropic’s entry was “a sign of intensifying competition” for legal tech incumbents.

The plugins are open-source and file-based, which means teams can inspect, modify, and extend them. One can build entirely custom plugins using a launch plugin called Plugin Create, and Anthropic says organization-wide sharing and private marketplaces are in development. Matt Piccolella, on the Anthropic product team, told TechCrunch that the goal was simply “bringing them to Cowork and giving them that kind of user-friendly, UI-centric flavor that will allow the maximum number of people to use them.”

The practical reality, at least for now, is uneven. Plugins currently save locally to individual machines, which limits their usefulness for teams trying to standardize workflows. Cowork itself is still a research preview with explicit warnings against regulated workloads. And the connectors only work as well as the underlying MCP integrations, which vary in maturity. But the architecture is the point: Anthropic is building a system where domain expertise is modular, shareable, and composable, not locked inside prompts that individual users have to reconstruct from scratch every session.

Apps Inside the Chat: Claude as Command Center

The third piece landed on January 26, when Anthropic announced that nine workplace tools were now available as interactive apps inside Claude’s interface. The launch partners were Slack, Canva, Figma, Box, Clay, Asana, Amplitude, Hex, and Monday.com, with Salesforce listed as coming soon.

This is not the same as Claude being able to “talk to” external services in the background, which it has been able to do through MCP connectors for some time. The difference is that the apps render interactive interfaces inside Claude itself. One can build an analytics chart in Amplitude and adjust parameters without leaving the conversation. One can draft a Slack message, preview the formatting, and review it before posting. One can turn a conversation into an Asana project with tasks and timelines that sync to the actual platform. The underlying technology is what Anthropic calls MCP Apps, an extension to the Model Context Protocol that lets any MCP server deliver an interactive user interface within any supporting AI product.

The strategic logic is transparent: the more work one can do without leaving Claude, the stickier Claude becomes. And because MCP is an open standard that Anthropic donated to the Linux Foundation’s Agentic AI Foundation late last year, the protocol benefits from adoption even by competitors. OpenAI adopted MCP across its products in March 2025 and launched its own Apps system in October.

There are notable gaps. Microsoft 365, Google Workspace, and several other suites that dominate corporate environments are absent from the launch lineup. The apps are not yet available inside Cowork, though Anthropic says that integration is coming soon, and one can imagine how powerful it would become when it arrives: an autonomous agent that can not only read and write local files but also update project boards, pull data from cloud storage, and post status updates to team channels, all within a single task flow. Mobile support was not mentioned.

What This Adds Up To

If one squints, the shape of Anthropic’s ambition becomes clear. Cowork is the execution layer, an agent that can do work autonomously on a user’s machine. Plugins are the knowledge layer, encoding domain expertise and workflow logic so the agent knows how to do specific jobs. And MCP Apps are the integration layer, connecting the agent to the broader ecosystem of tools where work actually lives. Stack all three, and what emerges is less a chatbot than a general-purpose work interface, one where the AI is not just answering questions but orchestrating tasks across files, applications, and services.

Whether this constitutes an existential threat to the SaaS industry, as the stock market’s initial reaction implied, is a separate question. Gartner analysts wrote that Cowork and its plugins are “potential disrupters for task-level knowledge work but are not a replacement for SaaS applications managing critical business operations.” That distinction matters. Cowork can draft a contract review memo; it cannot replace the case management system that tracks a law firm’s entire docket. It can reconcile expenses from receipt screenshots; it cannot serve as a company’s accounting system of record.

But the “just task-level” framing may be too generous to incumbents. A great deal of what people pay for in software subscriptions is precisely the task-level work: the data entry, the formatting, the cross-referencing, the drafting, the organizing. If an agent can do that work at $20 per month while also reading local files and connecting to the tools where the real data lives, the question for many SaaS companies is not whether their core platform is threatened but whether their per-seat pricing model survives contact with a world where the seat is occupied by an AI.

All of this remains early in ways that matter: Cowork is still a research preview with explicit caveats about destructive actions, the plugins save locally rather than syncing across teams, and the app integrations do not yet work inside the autonomous agent. But the architecture is coherent, the pieces are shipping at a pace that suggests Anthropic views the window for establishing this kind of platform position as narrow, and the market has already priced in the possibility that this trajectory will continue. For anyone whose work involves files, documents, and the tedious connective tissue between applications, the coming months will reveal whether what Anthropic has built is a genuine new layer of work infrastructure or an impressive demo that buckles under the weight of real organizational complexity.