MCP Hits 97M Installs. Claude Cowork Is Killing Legal SaaS.

Anthropic's Model Context Protocol crossed 97 million installs in April 2026. That number needs context: six months ago it had under 5 million. The jump is not viral adoption — it's the result of MCP being bundled into Claude Desktop, VS Code extensions, and a growing list of developer tools that have adopted it as the default way to connect AI assistants to external systems.

At the same time, Claude Cowork — Anthropic's workplace-integrated version of Claude — is being credited with what legal tech investors are calling a "SaaSpocalypse": a sustained revenue decline across contract review, document analysis, and billing management SaaS companies as their core workflows get absorbed into a single Claude interface.

These two stories are connected. MCP is why Claude Cowork can replace dedicated SaaS — and why the same disruption is coming to every category of workflow software.

What MCP Actually Is and Why 97M Matters

The Model Context Protocol is a standardised interface that lets AI models talk to external data sources and tools without custom API integration for each connection. The analogy in developer terms: MCP is to AI assistants what USB-C is to device connections. Before USB-C, every device had its own connector. After, one standard cable works everywhere.

Before MCP, connecting Claude (or any LLM) to your internal tools required:

• A custom API wrapper for each tool
• Prompt engineering to describe each tool's schema to the model
• Custom error handling for each integration
• Maintenance when the tool's API changed

With MCP, you write one MCP server for your tool, and any MCP-compatible client (Claude Desktop, VS Code Copilot, Cursor, Zed, dozens of others) can use it. The developer writes the integration once; the AI ecosystem consumes it everywhere.

97 million installs means the tooling ecosystem is already built. There are now MCP servers for GitHub, Linear, Notion, Slack, Jira, Salesforce, PostgreSQL, MongoDB, AWS services, and hundreds of internal enterprise tools. The question is no longer "can my AI assistant connect to this?" — it's "which AI client do I prefer for the work I'm doing?"

This is the threshold where a protocol stops being experimental and becomes infrastructure. MCP crossed it somewhere around February-March 2026, and the April install number confirms it.

Claude Cowork: What It Is and What It Does

Claude Cowork is Anthropic's enterprise product that integrates Claude into workplace environments with persistent context, team-shared memory, and workflow automation. It is not just "Claude with a company account" — it is a different product architecture where Claude maintains context across sessions, knows your company's documents and processes, and can initiate multi-step workflows across connected tools via MCP.

The specific capabilities that are eating legal SaaS:

Contract review: Claude Cowork reads contracts against a company's standard terms template, flags deviations, explains the legal implications of non-standard clauses, and drafts redline suggestions. This was the core product of ContractSafe, Ironclad, and similar contract lifecycle management (CLM) tools. Those tools charged $500–$2,000/month for the contract review workflow. Claude Cowork does it as a conversation.

Document analysis and due diligence: Legal teams doing M&A or litigation discovery were spending $20K–$100K on document review platforms (Relativity, Everlaw, Disco). Claude Cowork with a document corpus MCP server reads the same documents, answers questions, identifies relevant passages, and produces summaries without a separate platform subscription.

Billing and time management: Legal billing software (Clio, MyCase, TimeSolv) charges $49–$99/user/month for time tracking and invoice generation. Claude Cowork connected to a calendar and email MCP server can reconstruct billable time from meeting records and emails, then draft invoices. Not perfect, but good enough to reduce the number of billing software seats a firm needs.

The pattern: Claude Cowork is not replacing legal expertise. It is replacing the middleware layer between legal expertise and the documents and systems that expertise operates on. That middleware layer was worth $4B+ in annual SaaS revenue in legal tech alone.

The SaaSpocalypse: Which Categories Are Exposed

"SaaSpocalypse" is hyperbolic but the directional claim is accurate. The categories most exposed to Claude Cowork displacement share three characteristics:

1. The core workflow is reading and generating structured text
2. The main product value is workflow orchestration, not proprietary data
3. The switching cost is document migration, not deep technical integration

Legal tech is the first visible casualty because law firms are high-value, cost-sensitive, and almost entirely text-based in their work. The same disruption is visible in:

HR tech: Performance review software (Lattice, Leapsome, 15Five) that charges $6–$14/user/month for structured templates and review workflows. Claude Cowork with an HR MCP server does the same templates in conversation.

Customer success platforms: Tools like Gainsight and ChurnZero that structure customer health scores and playbooks. These are text-manipulation workflows on top of CRM data. MCP-connected Claude can read the CRM data and run the same playbooks without a separate $500K/year platform.

Proposal generation tools: Pandadoc, Proposify, RFPIO — all tools for generating proposals from templates and pulling in pricing data. A Claude with MCP access to your CRM, pricing tables, and proposal templates does the same work.

The pattern does not apply to tools with proprietary data moats (Bloomberg Terminal, Veeva in pharma), deep technical infrastructure integration (Salesforce CRM itself, not its proposals feature), or where the value is a marketplace rather than a workflow (LinkedIn, Upwork).

What This Means for MCP Server Development

The 97M install count is the signal that building MCP servers is now commercially valuable — not just technically interesting. The market for well-built MCP servers is:

Enterprise integration agencies: Companies are paying $50K–$200K to systems integrators to build MCP servers for their internal tools. The demand is real and current. If you have TypeScript or Python skills and understand API integration, building and selling MCP servers for popular enterprise tools is a real market.

Open source MCP server libraries: The most installed MCP servers on GitHub are seeing significant contribution activity. Building the definitive MCP server for Salesforce, for ServiceNow, for Workday, or for industry-specific tools (Epic for healthcare, Caseware for accounting) is a defensible open source position that drives significant traffic and potential commercial licensing.

MCP-native SaaS: The smart response to Claude Cowork disruption is not to fight Claude Cowork — it is to become an MCP server. A contract review SaaS that positions itself as "the best MCP server for contract management" is complementary to Claude rather than competitive with it. The value shifts from workflow to data structure and audit trail.

The Protocol Competition: MCP vs OpenAI's Tools

MCP has a potential competitor: OpenAI's own tool use specification, which predates MCP but has not been standardised as an open protocol. OpenAI has announced "Operator" capabilities that function similarly to MCP in connecting agents to external tools. The key difference:

• MCP is an open protocol — any client, any server, any model
• OpenAI's tool use is a proprietary API format — works best with OpenAI models, requires adaptation for other models

For developers building tool integrations today, MCP is the better bet for two reasons: cross-model compatibility (same server works with Claude, with Cursor's built-in model, with whatever Spud exposes) and existing ecosystem (97M installs of clients means your server has immediate potential reach).

If OpenAI open-sources its Operator tool format and achieves cross-model adoption, the calculus changes. Until then, MCP is the default protocol for agentic AI tool integration.

Key Takeaways

• MCP crossed 97M installs in April 2026 — transitioned from experimental protocol to infrastructure standard; ecosystem of MCP servers for GitHub, Linear, Slack, AWS, and 500+ other tools already exists
• Claude Cowork is displacing legal SaaS — contract review (CLM tools), document analysis (Relativity/Everlaw), and billing (Clio/MyCase) are the first visible casualties as the middleware layer between legal expertise and document systems gets absorbed into Claude
• The SaaSpocalypse pattern applies to any SaaS where core value is text-manipulation workflow, not proprietary data — HR performance reviews, customer success playbooks, and proposal generation are next
• MCP server development is now commercially valuable — $50K–$200K enterprise integration projects, open source positions in under-served verticals, and MCP-native SaaS pivots are all real opportunities
• MCP vs OpenAI tools: MCP is the open-protocol winner unless OpenAI standardises Operator — build for MCP now for cross-model compatibility
• The smart SaaS response to disruption: become an MCP server with data structure and audit trail value, not try to fight Claude on workflow orchestration

Track what Claude and competing models are costing developers right now with LLM API Pricing. For the broader AI job market impact, read where the tech jobs went after AI hiring wave. Compare Claude against GPT-5 for your specific use case with Claude vs ChatGPT.