The MCP market: why every AI vendor bet on one protocol

MCP moved faster than almost any integration standard in recent memory. Anthropic open-sourced it in November 2024; on 26 March 2025 OpenAI shipped support in its Agents SDK; on 21 May 2025 OpenAI added remote MCP servers to the Responses API and joined the steering committee; and in December 2025, thirteen months in, Anthropic donated the protocol to the Agentic AI Foundation under the Linux Foundation, co-founded by OpenAI and Block, with AWS, Google, Microsoft, Cloudflare, GitHub and Bloomberg as supporting members. This is the market picture: who adopted it, why now, and where it bites.

What the protocol is, in one breath

MCP is a JSON-RPC 2.0 protocol that lets an AI client (agent) discover and call typed capabilities on a server — tools, resources and prompts. Open standard, open source, speaking stdio or HTTP. If you want the hands-on version — building and publishing through one — read how-this-post-was-written on this same blog; this post is the map, not the territory.

The timeline that matters

• 25 Nov 2024 — Anthropic open-sources the Model Context Protocol.
• 26 Mar 2025 — OpenAI adds MCP support to its Agents SDK (v0.0.7).
• 10 Apr 2025 — n8n ships MCP Client Tool and MCP Server Trigger nodes (v1.88.0).
• 20 May 2025 — Google adds MCP support to the Gemini SDKs.
• 21 May 2025 — OpenAI ships remote MCP servers in the Responses API and joins the MCP steering committee.
• Dec 2025 — Anthropic donates MCP to the Agentic AI Foundation (Linux Foundation); OpenAI and Block co-founded; AWS, Google, Microsoft, Cloudflare, GitHub, Bloomberg supporting.
• Jan 2026 — MCP Apps ships as the first official extension.
• Mar 2026 — the 2026 roadmap lands with enterprise readiness as the top priority; AAIF lists 286 member organisations.
• Jul 2026 — the 2026-07-28 spec aligns MCP auth with OAuth 2.0 and OIDC.

The numbers — reported honestly

Published figures disagree, so treat any single one as directional. Here is what the sources say, attributed:

• SDK downloads — roughly 97M a month was reported in March 2026; npm's TypeScript SDK alone recorded 231.9M downloads in September 2026 (up from 31.3M a year earlier); the project puts its Tier-1 SDKs close to half a billion a month.
• Servers — 5,800+ published servers (March 2026 directory count); other directories put it at 10,000+ public servers and 17,000+ by Q1 2026. The real answer: thousands, growing weekly.
• Clients — 300+ client applications in the ecosystem by early 2026.
• Governance — 286 member organisations under the AAIF as of late 2026.
• Analyst estimates (quote with care) — one estimate puts Fortune 500 MCP adoption at a 28% implementation rate in under 18 months; these are projections, not audits.

Worth repeating: several widely shared MCP statistics have no traceable source. Directories disagree, download counts move by tens of millions a month, and self-reported numbers are the norm. The shape of the trend is reliable; the exact digits are not.

Who is actually on board

• AI platforms — Anthropic, OpenAI, Google DeepMind, Microsoft. All ship or plan MCP-first tool access.
• Cloud — AWS, Cloudflare, Azure. Cloudflare ships native remote-MCP hosting for its platform.
• Dev tools — GitHub / VS Code Agent Mode, Cursor, Replit, Sourcegraph, Zed, JetBrains.
• Enterprise software — Salesforce, Atlassian (Jira), Notion, Figma, Asana, Slack.
• Finance — Block and Bloomberg are both members of the foundation.
• Automation — n8n ships MCP client and server nodes; the workflow-automation crowd adopted it early.

Why now — the four reasons it is shipping this fast

• Vendor neutrality. The steering committee moved from Anthropic to the Linux Foundation in 13 months — procurement can finally sign off on one protocol without betting on one company.
• Function calling needed a substrate. OpenAI adopted MCP to give its function-calling infrastructure access to the ecosystem rather than maintain a parallel plugin architecture.
• Governed agent access is the enterprise ask. One connector per system, permissions and audit centralised — exactly the pattern (per-credential audit, read-only views, explicit delete consent) this site's MCP uses.
• Composability beats bespoke glue. Every team that wrote a custom tool-caller in 2023-24 paid the same tax twice; MCP is the shared layer that tax pays for.

Where it bites

• Quality is uneven. One April 2026 reliability study of 100 production MCP servers measured pass rate, error class, latency and tool-call success — and found wide variance. Read the server before you install it; it is instructions your agent will execute.
• Permissions are the risk. MCP amplifies whatever you grant an agent — an over-scoped server is an open door, not a features page.
• Auth kept evolving. The 2026-07-28 spec aligned MCP auth with OAuth 2.0 and OIDC for exactly this reason; earlier servers bolt auth on in awkward shapes.
• Supply-chain exposure. Public server registries lower the bar for typosquatting and malicious servers — review like you review npm dependencies.
• Hype metrics everywhere. If a number sounds precise, ask whose dashboard it came from.

What I would do with this

• Start with one read-mostly server and instrument it — logs, latency, error budget — before giving an agent write access.
• Put audit in from day one (who/what/when), not after the first incident.
• Prefer official SDKs (TypeScript or Python) and pin versions — the ecosystem moves fast enough to break you.
• If you already run a product, consider exposing one governed MCP server instead of five ad-hoc integrations — the compounding win is real.
• If you are evaluating agentic products, ask each vendor how their MCP story works — their answer tells you more than their demo.

Spec at modelcontextprotocol.io; governance via the Agentic AI Foundation (a Linux Foundation foundation). This site runs a small production example — an OAuth-2.1-protected MCP server that publishes this very journal; the companion post, how-this-post-was-written, shows that build from the inside.