Companies ahead in the game

What “ahead” means here

A company is ahead, on this reading, when it has produced something other people can build on with knowledge: an open standard someone else implemented, a documented interface an agent engine can rely on, or a product surface that grounds its answers in a specific corpus.

Model quality is a different axis, and a faster-moving one. The corpus behind this page records model releases as dated entries in a chronology, not as evidence of advantage. The negative case is worth stating plainly: Safe Superintelligence, founded in June 2024 by Ilya Sutskever with Daniel Gross and Daniel Levy, is recorded as having no product line and no revenue, with roughly twenty employees at the time of its March 2025 round and a valuation of 30 billion dollars. That is a fact about capital and reputation, and it is excluded from the groups below.

Who is visibly doing what

Anthropic: a standard that left the building

Anthropic originated the Model Context Protocol, released on 25 November 2024 and created by its engineers David Soria Parra and Justin Spahr-Summers, and donated it in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation that Anthropic, Block and OpenAI co-founded. Agent Skills, published the same month, is the second contributed standard. On the product side, Claude reached a tenant's documents through Microsoft 365 Copilot from 9 March 2026, Claude Cowork shipped an enterprise release in February 2026, and Claude Code became generally available in May 2025.

OpenAI: published interfaces, and support rather than a rival protocol

OpenAI adopted MCP in March 2025 and documents MCP connections inside its own agent tooling, alongside an Agents API, sandboxes, multi-agent orchestration, guardrails and evals — documented surfaces rather than demonstrations. The corpus records that choice as support for an existing standard rather than the publication of a competing one. On the enterprise side, ChatGPT Enterprise and Team add organisational separation and administrative controls, and the product line includes grounding on an enterprise corpus. GPT-5 is described not as a single network but as a system with a fast model, a deeper reasoning model and a router that chooses between them.

Google and Google DeepMind: the substrate, and grounding inside the document surface

The Transformer, published by Google researchers in 2017, is the largest single contribution attributable to Google in the corpus: the architecture the foundation-model domain is built on. In the enterprise knowledge use case, Gemini for Workspace competes with Copilot on the same document surface, and Gemini Notebook (formerly NotebookLM) performs grounded retrieval over uploaded documents. Gemini 1.5 shipped with a one-million-token context window in February 2024. The corpus also records a distinction it treats as material: Google is an adopter of MCP rather than its author, and the graph keeps an author edge and an adopter edge apart.

Replit: a grounded product surface, shipped to non-developers

Replit Agent runs on Claude — Sonnet 4.6 for sustained development work, Opus 4.7 for architecture and multi-file refactoring — and is served through Google Cloud's agent platform. Agent sessions run for six hours and longer, and the same case study describes a meta-agent that builds other agents. Replit adopted MCP, and its buyer is explicitly not only the professional developer: students, teachers, designers and small business owners are named as the target set.

What is verifiable versus vendor-reported

The corpus is explicit about the difference, and this page keeps it. For enterprise knowledge, almost all product claims are vendor-described; the most independently attested items are the Bing Chat launch and roll-out dates, the Prometheus model built on OpenAI's GPT line, and Perplexity's 2023 legal and scraping controversies. Headline productivity numbers are marketing and are recorded only as a product's own description.

MCP's adoption metrics — more than 10,000 production servers and about 97 million SDK downloads a month — come from the protocol's maintainers. Anthropic's $65B Series H at a $965B post-money valuation on 28 May 2026 is company-announced and separately reported; its February 2026 round figures come from analyst trackers rather than an announcement. Run-rate revenue crossing $47B is company-reported and unaudited, and headcount is not disclosed, with third-party estimates differing by a factor of two. Replit's revenue is recorded as three conflicting claims rather than one figure.

What a reader should copy first

Publish an interface, not a connector. MCP's spread came from a documented surface other vendors could implement; the same logic applies to a site that wants to be usable by agent engines — see llms.txt and the agent-to-agent surface in A2A.

Ground the product on a named corpus, and say which one. Enterprise knowledge is the use case where grounding is on private, enterprise-held data, which is why retrieval and access control are its central capabilities — see Enterprise and knowledge graphs.

Keep the evidence label with the claim, and start with the knowledge layer rather than the model, because the layers worth copying are the ones that stay useful when the model line changes. Adoption covers how that is put into practice.