Anthropic's Model Context Protocol surpasses 97 million installs

When Anthropic released the Model Context Protocol in late 2024, the reception was cautiously optimistic. Developers saw promise in a standardized way for AI agents to talk to external tools. Eighteen months later, MCP crossed 97 million installs in March 2026, and the conversation has shifted from 'should we adopt this?' to 'how do we build on top of it?'

The number itself tells part of the story. For context, npm packages considered widely adopted in the JavaScript ecosystem typically sit in the low tens of millions of weekly downloads after years of use. MCP reaching 97 million total installs across a much shorter window puts it in rare company for a protocol that touches infrastructure rather than end-user software.

What MCP actually does

At its core, MCP is a specification for how AI agents connect to external systems. Before it existed, each agent framework had its own way of calling APIs, reading databases, or triggering workflows. That fragmentation meant a tool built for LangChain would not work with an AutoGen agent without significant rewiring. MCP standardizes the handshake. An agent that speaks MCP can, in theory, connect to any MCP-compatible server without custom integration code.

Think of it the way HTTP standardized how browsers and servers communicate. HTTP did not invent the web, but it made the web possible at scale. MCP is attempting something similar for the layer between AI reasoning and the real-world tools those models need to act on.

Every major AI provider now ships MCP support

What changed the trajectory of MCP was not just Anthropic's push. OpenAI, Google DeepMind, Mistral, and several other AI providers added MCP-compatible tooling to their agent SDKs over the course of 2025. That cross-vendor adoption removed the primary objection developers had: betting on a protocol controlled by a single company. Once the major players aligned on MCP as the default mechanism, third-party tool builders had a clear target.

GitHub now lists over 12,000 public MCP server repositories. These range from servers that expose Postgres databases to AI agents, to ones that wrap Stripe, Notion, Salesforce, and dozens of other SaaS platforms. A developer building an AI assistant for financial operations can wire it to live data sources in hours rather than days.

Why developers chose MCP over alternatives

Several competing standards were proposed between 2023 and 2025. Most failed to gain traction because they were too tightly coupled to specific frameworks or required heavy client-side configuration. MCP won on simplicity. The server-side implementation is straightforward enough that a single developer can expose a new tool in an afternoon. The client-side consumption is handled by the agent framework, so application developers rarely need to think about the protocol at all.

There is also a security model baked into the spec. MCP servers declare their capabilities upfront, and clients can enforce permissions before any tool call is made. For enterprise teams worried about agents accessing systems they should not, that explicit permission layer was a practical requirement, not a nice-to-have.

What 97 million installs means for the ecosystem

Reaching this install count does not mean every deployment is production-grade. A significant share will be developer experiments, tutorials, and internal proofs of concept. That is normal for any protocol in its early scaling phase. The more telling signal is that enterprise software vendors are now shipping MCP support as a default feature rather than a beta add-on. When a product manager at a mid-size company can connect their internal AI assistant to their CRM by toggling an MCP setting, the protocol has moved past the developer-only phase.

Anthropic has not published a breakdown of installs by industry or geography, but activity on the MCP GitHub repository and community forums suggests strong adoption in fintech, legal tech, and developer tooling companies. These are sectors where agents need reliable, auditable access to structured data, exactly the problem MCP was built to address.

What comes next for MCP

The spec is not frozen. Anthropic and the broader working group have been iterating on streaming responses, multi-agent communication patterns, and improved authentication flows. Version 0.9 of the spec, expected in Q2 2026, is supposed to address long-running tool calls, which is one of the more common pain points developers hit when building agents that interact with slow external systems like legacy ERP platforms.

For now, the 97 million install figure is concrete evidence that the fragmented, framework-specific approach to agent tooling is giving way to something more uniform. Whether MCP holds that position or gets displaced by a later standard is an open question, but as of March 2026, it is the protocol that the majority of the AI agent ecosystem is building on.