MCP didn’t win by accident.
Anthropic’s confirmation that the Model Context Protocol has reached 97 million monthly SDK downloads, verified from Anthropic’s official announcement page, marks something more significant than a metric milestone. It marks the point at which a protocol becomes infrastructure. The comparison point is HTTP in the early web era: once enough systems spoke the same protocol, the network effects compounded. Anthropic reports more than 16,000 active MCP servers, spanning developer tooling through Fortune 500 enterprise deployments, though the precise figure remains directionally confirmed rather than independently audited.
Why does this matter for legal AI specifically? Because the execution-layer promise, AI that acts in legal workflows rather than recommends, depends entirely on integration depth. An agent that can reason about a contract but can’t pull the signed version from the document management system, can’t check the client matter in the billing platform, and can’t log the action in the case management tool isn’t an execution layer. It’s a smart chatbot with good legal knowledge. MCP is the connective tissue that makes the difference between those two things.
What Claude for Legal Actually Is
Claude for Legal reached general availability on May 15, configured from the Claude 4 family for legal workflow execution. Per Anthropic sources, the product integrates with legal tools through MCP connectors and is positioned at enterprise legal teams and law firms with contract lifecycle and legal operations workflows as primary targets.
The “execution layer” framing that’s circulated in coverage this week is observer characterization, not Anthropic’s language. A distinction Anthropic’s technical webinar appeared to reinforce, but treating it as an Anthropic commitment rather than an analyst framing would be a misread of what’s confirmed. What is confirmed: Claude for Legal operates on agentic principles, uses MCP for tool-use, and is designed for action, not just analysis.
Pricing is a gap in the current record. Reportedly, the product includes separate programmatic use budgets billed at full API rates for enterprise subscribers, but the primary source for that claim is a broken URL. Don’t model your legal operations budget around a figure from a broken source. Verify directly with Anthropic before any procurement conversation. This is also a distinct product from the Claude for Small Business launch on May 16, which covered the enterprise AI spend trajectory with 20+ legal MCP connectors for DocuSign and LexisNexis. Same underlying model family, very different buyer profile and commercial structure.
The Competitive Map
Three products are now actively competing for enterprise legal AI budget.
Anthropic, Claude for Legal. Architecture: MCP-native, agentic, built for execution over recommendation. Differentiator: the 97M-download MCP ecosystem means the integration surface theoretically extends anywhere a developer has built a server. The catch: 16,000 servers skews heavily toward developer tooling. The legal-specific MCP server ecosystem, contract management platforms, e-discovery tools, court filing systems, matter management software, is substantially smaller. The integration breadth promise is only as good as the legal-vertical server coverage, and that’s a due diligence question, not a launch-day given.
Microsoft, Legal Agent for Word. Architecture: deeply embedded in Office and Microsoft 365, with MCP integration into iManage and NetDocuments, per the registry brief from May 5. Differentiator: the existing install base. Enterprise law firms and corporate legal departments that have already standardized on Office aren’t evaluating an AI tool, they’re evaluating a feature upgrade to infrastructure they already run. The friction of adoption is near zero if you’re already a Microsoft shop. The limitation: it’s a Word-centric product by design. For legal operations workflows that span beyond document drafting, matter intake, billing review, compliance monitoring, a Word-embedded agent has structural limits.
Harvey. Architecture: legal-domain-specific training on large volumes of legal text, case law, and contract data. Differentiator: domain depth. Harvey’s position is that general-purpose LLMs, even well-configured ones, don’t have the same baseline legal knowledge as a model trained specifically on legal corpora. The competitive risk Harvey faces is capability convergence, as Claude 4, GPT-5, and Gemini Ultra expand their training data and legal fine-tuning, the domain knowledge gap narrows. Harvey’s durability depends on whether domain-specific training remains a meaningful differentiator as frontier model capability increases.
Clio deserves a separate mention as context. The $500M funding round Clio closed positions it as the legal practice management platform with the most AI integration surface in the mid-market legal segment. Clio’s MCP server availability, or lack of it at launch, will be a meaningful signal for Claude for Legal’s practical reach into the market segment below Am Law 200 law firms.
MCP as Governance Vector
Enterprise legal teams evaluating any of these products should read why agentic AI is harder to certify under the EU AI Act before a deployment conversation.
Execution-layer AI in legal contexts creates specific governance questions that general enterprise AI deployment frameworks don’t address:
Attorney-client privilege: when an MCP-connected agent reads a privileged document to complete a task, what are the privilege implications? Most legal ethics guidance hasn’t caught up with agentic tool-use at this level of integration.
Authorization scope: MCP’s tool-use model requires explicit authorization for each connected server. In a legal context, that means defining, in advance, which systems the agent can read, which it can write to, and which actions it can take without human review. That’s not a configuration question. It’s a legal operations governance question.
Data residency: enterprise legal clients have data residency and sovereignty requirements. Claude for Legal’s MCP integrations touch data wherever the connected servers sit. The governance architecture needs to map those data flows before deployment.
None of this makes Claude for Legal unusable. It makes it a product that requires governance architecture work before deployment, work that’s entirely feasible but that doesn’t come in the box.
Procurement Considerations
The buyer who picks wrong between these three options doesn’t just pay the wrong price. They spend 18 months with an AI tool that integrates with the wrong systems, trains staff on the wrong workflows, and creates governance gaps that surface in the next audit cycle.
The evaluation framework that actually matters, in order:
One: map your integration requirements first. Which systems does your legal operation actually touch, document management, billing, matter management, court filing, compliance monitoring? Which of those systems have MCP servers (Anthropic ecosystem), Microsoft 365 integrations (Microsoft), or Harvey-native integrations? The product with the best integration coverage for your specific stack wins, regardless of benchmark performance.
Two: governance before capability. What are your data handling requirements? Where does client data need to stay? What actions require human review? Define this before evaluating any product.
Three: verify pricing under your actual usage pattern. Programmatic budget pricing at full API rates compounds fast at enterprise legal scale. Model it before commitment.
The MCP 97M downloads number is real. The execution-layer promise is directionally credible. The legal-specific integration ecosystem is still being built. Buy for where it’s going, but verify where it is today.