The last connector
On November 25, 2024, two Anthropic engineers — David Soria Parra and Justin Spahr-Summers — published an open specification to GitHub. The accompanying blog post opened with a problem statement the industry had been living with for two years: every team building AI applications was writing the same connectors, over and over — Postgres this week, Slack next week, GitHub after that — each one a custom integration, each one slightly broken in its own way. They called this the M×N problem. M models multiplied by N tools meant M×N integrations to build. The specification was their answer: the Model Context Protocol, or MCP.
The M×N framing had a history. In the 1970s, the same combinatorial logic had threatened to make computer networking permanently fragmented — every vendor’s equipment incompatible with every other’s. ISO’s response was the seven-layer OSI model: define clean interfaces, and the wall of bespoke drivers collapses. MCP applied the same logic to AI agents. Instead of each model learning the specific API of each tool, define one protocol: a server describes its capabilities in plain language; a client embedded in the AI application reads them; the model decides at runtime which to invoke.
The architecture was minimal by design. MCP divided the world into three roles: the host (the AI application itself), clients within it, and servers that exposed tools and data. Wire format: JSON-RPC 2.0. Authorization: standard OAuth. Nothing exotic. Anthropic shipped day-one servers for Google Drive, Slack, GitHub, Postgres, and Puppeteer — an implicit acknowledgment of which five integrations every team was already rebuilding from scratch.
Block, the payments company, deployed sixty internal MCP servers within months of launch. Developer tools — Zed, Replit, Codeium, Sourcegraph — adopted it before 2024 was out. Dhanji R. Prasanna, Block’s chief technology officer, put the appeal plainly: “Open technologies like the Model Context Protocol are the bridges that connect AI to real-world applications.” That kind of endorsement circulates in engineering Slack channels faster than analyst reports.
The signal came on March 26, 2025, when OpenAI announced MCP support across its products — the same day Anthropic released a specification update adding OAuth 2.1 and streamable HTTP transport. Google DeepMind confirmed Gemini support the following month. A protocol released by a single company had become, inside four months, something that competing labs agreed to implement rather than replace. In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, with OpenAI, Google, Microsoft, and AWS as platinum members — the same companies building incompatible tool-calling systems twelve months earlier now sharing custody of a neutral standard.
By mid-2026, the ecosystem counted more than ten thousand active servers and 97 million monthly SDK downloads. Neither number was really the point. The point was that the question “which connector do we build for this model?” had gone from a sprint item to a solved problem — the kind of solved problem that disappears so thoroughly you stop noticing it was ever a problem at all.
The M×N grid didn’t vanish. It became somebody else’s protocol.
Sources
- Introducing the Model Context Protocol — Anthropic — November 25, 2024 announcement; David Soria Parra and Justin Spahr-Summers as creators; M×N framing; day-one servers; early adopters Block, Zed, Replit, Codeium, Sourcegraph; Dhanji R. Prasanna quote
- Model Context Protocol — Wikipedia — technical architecture; March 2025 OpenAI adoption; April 2025 Google DeepMind; December 2025 Linux Foundation donation; 10,000+ servers and 97M monthly downloads by mid-2026
- One Year of Model Context Protocol — Ajeet Raina — Block’s 60+ internal MCP servers; March 26, 2025 OpenAI announcement timing; growth trajectory