# The Interop Ledger **Last verified: July 2026.** We keep the chapters in this part at patterns and ratios, so this sheet is the one place the volatile agent-protocol facts live. When it disagrees with a spec page or a foundation announcement, the spec page wins. ## Model Context Protocol (MCP) MCP won the tool-connection layer and is now community property. Anthropic donated it in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded with Block and OpenAI. The spec ships as dated releases roughly twice a year. - **Current spec:** 2025-11-25. A release candidate for 2026-07-28 is out, the largest revision since launch: a stateless HTTP core plus extensions for server-rendered UI and long-running tasks. - **Adoption:** SDK downloads run on the order of 100 million a month, and every major assistant and coding tool ships a client (Claude, ChatGPT, Gemini, Copilot, Cursor, VS Code). - **Registry:** the official registry holds a couple of thousand entries; community indexes count well over ten thousand servers, the softest numbers on this sheet. The protocol and its client support are stable; the registry, enterprise governance, and the new extensions are still churning. ## Agent-to-agent protocol (A2A) A2A covers agents calling other agents across vendors. Google donated it to the Linux Foundation in 2025, and it reached 1.0 within its first year. - **Current version:** 1.0.0 per the spec site. - **Steward:** the Linux Foundation, with a technical steering committee drawn from AWS, Cisco, Google, IBM, Microsoft, Salesforce, SAP, and ServiceNow. - **Adoption:** over 150 supporting organizations, SDKs in Python, JavaScript, Java, Go, and .NET, and runtime support in Azure AI Foundry, Copilot Studio, and Amazon Bedrock AgentCore. The core spec and governance are stable; production evidence is thinner, mostly pilots inside the big platforms rather than open cross-vendor traffic. ## Agent payments No payment protocol has won yet. The contenders sit at different layers of the stack, and most production flows will touch more than one. - **AP2** (Google's Agent Payments Protocol): an open spec with over 60 partners, at version 0.2 as of spring 2026. It handles one-off purchases; recurring and peer-to-peer payments remain future work. Live pilots run through PayPal and Mastercard. - **ACP** (OpenAI and Stripe's Agentic Commerce Protocol): the spec is still maintained, but OpenAI retired Instant Checkout in March 2026 after weak merchant uptake, moving to retailer-run apps inside ChatGPT. The pattern that held is discovery in the assistant, purchase on the merchant's site. - **x402** (Coinbase-led, HTTP 402 plus stablecoins): over a hundred million on-chain transactions but only tens of thousands of dollars in genuine daily volume, much of it testing. Its foundation counts Cloudflare, Google, and Visa among members. - **Card networks:** Visa Intelligent Commerce and Mastercard Agent Pay both issue scoped agent tokens, a card credential bound to one agent, one merchant scope, and one consent policy. The first live authenticated agent transactions ran in Asia in spring 2026. The card-network token model looks stable; everything above it, including which protocol merchants will actually integrate, is still churning. ## Web bot authentication The web is moving from guessing a bot's identity by IP address to verifying it with cryptography. - **Web Bot Auth:** agents sign requests with HTTP Message Signatures (RFC 9421) and publish public keys at a well-known URL. The IETF chartered a working group in 2026 to standardize it. - **Deployment:** Cloudflare verifies these signatures in production, and AWS WAF, Akamai, Vercel, and Shopify have shipped support. - **Paid access:** Cloudflare's pay-per-crawl beta charged per fetch and answered non-payers with HTTP 402. A July 2026 successor pays publishers when their content appears in an AI answer, and from September 2026 mixed-use crawlers are blocked on ad-supported pages by default. The signature mechanism is stable; the business model on top of it changes quarterly. ## Agent-readable docs (llms.txt) llms.txt is a convention, not a standard, and the measured reality is thin. Roughly one in ten sites in large 2026 crawls carry the file, concentrated in developer-tool and documentation sites, and the major AI crawlers almost never fetch it: a 90-day sample of half a billion AI bot visits logged a few hundred fetches. Google has said it will not support the convention. Publish one if coding agents consume your docs; do not expect it to change how assistants cite you. ## Agent traffic measurements Cloudflare Radar publishes crawl-to-referral ratios, meaning pages fetched per referred visitor. Early-2026 figures put GPTBot above 1,000 crawls per referral and ClaudeBot above 20,000, while traditional search engines sit near single digits. By mid-2026, under 10 percent of AI crawler requests were search-purpose crawling that could return a citation; the rest fed training and answer generation. The exact ratios move every quarter; the imbalance has held for a year. ## How to read this sheet Protocol facts rot in months. Before you rely on any line here, check this sheet's date and the primary source. A version number more than two quarters old is a hypothesis, not a fact, and adoption counts age even faster. The stable-versus-churning judgments are the durable part; reverify everything else before it goes into a roadmap or a build decision. ## Where to reverify The MCP spec and blog at modelcontextprotocol.io, the A2A spec at a2a-protocol.org, the AP2 docs at ap2-protocol.org, the ACP repository from OpenAI and Stripe, the x402 docs on the Coinbase Developer Platform, the IETF Web Bot Auth working group on the datatracker, and Cloudflare Radar's AI insights pages for current traffic ratios.