Model Context Protocol

source

public(spec schema and docs in one repo plus reference SDKs in separate public repos)

license

Apache-2.0(OSI, new code/spec)+MIT(OSI, legacy contributions)

docs

CC-BY-4.0(documentation other than the specifications

governance

Linux-Foundation/AAIF(open, co-founded Anthropic/Block/OpenAI)

core-gated

ungated(spec schema and docs ship in the one LF/AAIF-governed repo and every reference SDK is a public modelcontextprotocol repo - there is no vendor tier to withhold anything for)

An open protocol with OSI-licensed reference implementations under open foundation governance, and nothing in the core held back. The specification, the JSON/TypeScript schema and the docs all live in one repository, every reference SDK is a separate public modelcontextprotocol repo (mcp-python-sdk and mcp-typescript-sdk are both on this map at 5), and the Linux Foundation's AAIF governs the project, so there is no vendor tier for anything to be withheld for. The LICENSE splits three ways: Apache-2.0 for new code and specifications, CC-BY-4.0 for documentation other than the specifications, and MIT retained for pre-transition contributions whose authors did not consent to relicensing. Only the two code licenses bear on the score. A compound license is resolved on every part and takes the most restrictive of them, so filing the documentation license alongside the code would let a license covering the project's prose govern the artifact you actually run and drag the product down to source-available. The documentation license is therefore recorded apart from the code licenses, the same way autogen records the identical situation - MIT code, CC-BY-4.0 docs, scored 5. Both code licenses are OSI-approved, so the tier holds.

• https://github.com/modelcontextprotocol/modelcontextprotocol/blob/main/LICENSE recorded 2026-06-04

  Apache-2.0 (new code/spec) + MIT (legacy) + CC-BY-4.0 (docs) license text

• https://en.wikipedia.org/wiki/Model_Context_Protocol recorded 2026-06-04

  Anthropic origin; donated to Linux Foundation AAIF Dec 2025

• https://raw.githubusercontent.com/modelcontextprotocol/modelcontextprotocol/main/LICENSE recorded 2026-08-12

  The three-way split, read rather than assumed. New code and specifications are under the Apache License, Version 2.0; documentation other than the specifications is under CC-BY-4.0; contributions made before the transition whose authors did not give relicensing consent remain under the MIT License, with "no rights beyond those granted by the applicable original license" conveyed. Copyright (c) 2024-2025 Model Context Protocol a Series of LF Projects, LLC.

• https://api.github.com/repos/modelcontextprotocol/modelcontextprotocol recorded 2026-08-12

  Repo metadata for the spec repo. It holds the specification, the protocol schema (TypeScript, emitted as JSON Schema) and the documentation; the reference SDKs are separate repos in the same public org. Nothing is withheld from it and there is no second, differently-licensed tree.

De facto standard agent tool protocol, adopted natively by OpenAI (ChatGPT, Agents SDK, Responses API), Google DeepMind (Gemini), Microsoft and AWS, with more than 16,000 MCP servers in the wild. The band rests on the two reference SDK registries - 335,363,270 PyPI downloads in the trailing 30 days for `mcp` and 198,046,096 npm downloads for @modelcontextprotocol/sdk - either of which clears the >10M level-5 floor by more than an order of magnitude.

• https://en.wikipedia.org/wiki/Model_Context_Protocol recorded 2026-08-13

  Adoption by OpenAI, Google DeepMind, Microsoft and AWS; reference SDKs across seven-plus languages; donation to the Linux Foundation.

• https://zuplo.com/blog/one-year-of-mcp recorded 2026-08-13

  Dated 2025-11-24. "Over 16,000 MCP servers exist in the wild. OpenAI has adopted it across ChatGPT (sort of), their Agents SDK, and their Responses API. Google DeepMind announced MCP support for Gemini. Companies like GitHub, Block, Apollo, Replit, Sourcegraph, Linear, Zapier and many others have integrated it into their platforms." NOT on this page, contrary to what this source previously claimed: any SDK download figure, the number 97M, or a 4,750% growth rate.

• https://pypistats.org/api/packages/mcp/recent recorded 2026-08-13

  last_month = 335,363,270 downloads for the reference Python SDK package `mcp`.

• https://api.npmjs.org/downloads/point/last-month/@modelcontextprotocol/sdk recorded 2026-08-13

  198,046,096 downloads for the window 2026-07-11 to 2026-08-09 for the reference TypeScript SDK.

A wire protocol is not benchmarked, so capability is left unscored here by policy rather than for want of a number: openness and adoption carry the whole signal for a specification. Two capability questions could be asked of a protocol and neither survives scrutiny. Breadth of implementation is adoption wearing a capability label, and this product's adoption is already banded on exactly that kind of evidence - 335 million monthly PyPI downloads of `mcp` and 198 million monthly npm downloads of @modelcontextprotocol/sdk - so scoring breadth as well would let one instrument vote twice and average adoption against itself. Spec maturity is genuinely independent of uptake, but no rubric on this map has a ladder for it, it would be a fifth instrument sharing a field with benchmark, feature-comparison and internal-evaluation scores, and with only two protocols on the map there is nothing to anchor a comparison against. A two-product feature comparison against the Agent2Agent protocol is the only thing the existing machinery could enforce, and it would require scoring that record's capability first. No score is invented to close the gap. The blank should therefore be read precisely: not "nobody has got to this yet" but "there is no independent capability question to ask of a specification". One cost comes with it, stated rather than papered over - with no capability score the maturity calculation falls through to adoption alone rather than dropping the product, so this record computes a maturity of 5.0 from a single axis. That does not distort the category, which holds seven mature open products, six of them reaching 4.5 with no blank in the arithmetic. The Agent2Agent protocol has the same shape and carries the same policy.

• https://api.github.com/repos/modelcontextprotocol/modelcontextprotocol recorded 2026-08-14

  Repo metadata - description "Specification and documentation for the Model Context Protocol", topics docs / mcp / specification / standard, 8,950 stars, homepage modelcontextprotocol.io. What the project ships is a specification and its schema, and no benchmark result, leaderboard entry or performance figure is published for it.