MCP Adoption Statistics 2026: 435M Monthly Downloads, 10,000+ Servers, and Why the Control Layer Is Now Urgent

_Last updated: 2026-06-10_

When Anthropic donated MCP (Model Context Protocol) to the Linux Foundation on December 9, 2025, it cited 97 million-plus monthly SDK downloads. Six months later, registry data puts the official Python and TypeScript packages at roughly 435 million downloads a month combined, more than 10,000 active public MCP servers are registered, and the protocol is no longer an Anthropic project. It's a Linux Foundation standard. That adoption curve means transaction volume is arriving. The control layer for those transactions is behind.

---

Key takeaways

---

What is MCP and why does its adoption matter?

MCP (Model Context Protocol) is an open protocol that gives AI models a standardized interface to discover and invoke external tools, APIs, and data sources. Before MCP, every AI-to-tool connection was a bespoke integration. With it, any MCP-compatible model can call any MCP-compatible server without custom glue code.

When a model uses MCP to call a payment tool, a travel booking API, or a procurement system, the downstream action is a real transaction with real money attached. So MCP adoption statistics are a leading indicator of agentic transaction volume. More servers means more tools, and more tools means more things agents can do with money.

For a precise breakdown of what MCP does and doesn't do in the payments context, see [Introducing the Shatale MCP Server](/blog/introducing-shatale-mcp-server).

---

What are the MCP adoption numbers for 2026?

| Metric | Figure | Notes |

|---|---|---|

| Monthly Python SDK downloads (mcp on PyPI) | ~283M | PyPI Stats, June 2026 |

| Monthly TypeScript SDK downloads (@modelcontextprotocol/sdk on npm) | ~155M | npm registry data, June 2026 |

| Combined monthly SDK downloads | ~435M | Python + TypeScript, June 2026 |

| Monthly downloads cited at donation announcement | 97M+ | Anthropic figure, December 2025 |

| Active public MCP servers | 10,000+ | Registered in public directories |

| Claude connectors at Linux Foundation donation | 75+ | At time of donation, December 2025 |

| MCP governance body | Agentic AI Foundation | Linux Foundation project |

| Co-founding members | Anthropic, Block, OpenAI | Announced December 2025 |

These numbers are estimates based on publicly available package repository data. Exact figures vary by measurement window and counting methodology, and registry downloads overcount humans (CI systems and mirrors pull packages too).

For context: react, the most-downloaded UI library on npm, pulls roughly 134 million downloads a week, around 580 million a month, as of June 2026. MCP's SDKs reached about three-quarters of that volume roughly 19 months after the protocol launched in November 2024. That rate reflects developers building production agent systems, not just experimenting.

---

What did the Linux Foundation donation change?

In December 2025, Anthropic donated MCP to the Agentic AI Foundation, a new Linux Foundation project. This wasn't a routine open-source formality. It moved MCP's governance from "Anthropic's protocol" to "neutral industry standard," and that shift carries real adoption weight.

Three things changed.

Start with vendor risk. Enterprises building on MCP had a reasonable concern that Anthropic could change direction, deprecate versions, or favor Claude. Under neutral Linux Foundation governance, MCP becomes infrastructure you can depend on no matter which model wins the market.

Then the competitive network effect. OpenAI and Block joining as co-founding members meant their platforms would support or align with MCP, expanding the pool of servers independent of Anthropic's own efforts.

And regulatory legibility. European enterprise buyers in particular are more comfortable adopting a protocol under Linux Foundation governance than one controlled by a single US commercial entity. That matters for enterprise deals in regulated industries.

The practical result: the 75+ Claude connectors at donation time represent early-adopter builders. The 10,000+ server count by mid-2026 represents mainstream developer uptake.

---

What does this growth curve imply for transaction volume?

At over 400 million monthly SDK downloads, a meaningful share of integrations connect to commerce-capable endpoints: marketplaces, travel platforms, procurement systems, payment APIs. Nobody publishes a count of how many MCP tool calls end in a charge. But even a small fraction of that activity is real money moving, and it's compounding faster than the control infrastructure around it.

The gap is specific. MCP standardizes how agents discover and call tools. It says nothing about whether an agent should be allowed to execute a given transaction, what policy applies, who approved the budget, or how the transaction is audited. Those questions belong to the control plane, which MCP explicitly doesn't govern.

That's not a flaw in MCP's design — keeping the data plane lean and the control layer separate is architecturally correct. But it means adopting MCP without a corresponding control layer is adoption without governance. The gap widens as volume grows.

---

Why can't control infrastructure wait?

The argument for waiting goes: "we'll add controls once we have scale." It's backwards.

Controls are hardest to retrofit at scale. When you have 10 agents transacting, an out-of-scope purchase is an incident. When you have 10,000 agents transacting, an out-of-scope rate of 0.1% is 10 incidents per day, each needing investigation, a potential dispute, and remediation.

Companies that build control infrastructure now, before their agent transaction volume is large, will have debugged, audited, institutionally approved systems in place when scale arrives. They'll also have the compliance posture (SOC 2 scope, policy documentation, audit logs) that enterprise buyers require as a condition of deployment.

The companies that wait will build their control layer under pressure, with live customer money at risk, in front of finance and legal teams that weren't in the room when the original deployment decision was made.

---

Frequently asked questions

How many MCP servers exist in 2026?

More than 10,000 active public MCP servers are registered in public directories as of mid-2026. The count of private and enterprise-deployed servers is higher but not publicly tracked.

How many times is the MCP SDK downloaded per month?

Roughly 435 million times per month combined as of June 2026: ~283 million for the Python mcp package (PyPI Stats) and ~155 million for @modelcontextprotocol/sdk on npm. Anthropic cited 97 million-plus monthly downloads at the December 2025 donation announcement. These are estimates; exact counts vary by measurement period.

Who governs MCP now that Anthropic donated it?

The Agentic AI Foundation, a Linux Foundation project. Anthropic donated MCP in December 2025; co-founding members include Anthropic, Block, and OpenAI.

Does MCP adoption mean more AI agent transactions?

Directionally, yes. MCP adoption tracks developers building agent systems, many of which connect to commerce-capable APIs. The control layer governing those invocations needs to keep pace with adoption.

Why doesn't MCP include payment controls?

By design. MCP is a data-plane protocol for tool discovery and invocation. Adding payment controls at the protocol layer would make it heavier and less broadly useful. Payment controls belong in a dedicated control plane that sits above payment execution, not inside the tool-invocation protocol.

---

The next piece worth reading is the Shatale MCP Server post. It covers how a control-plane MCP server gives your agents structured policy enforcement, not just payment execution: [Introducing the Shatale MCP Server](/blog/introducing-shatale-mcp-server).