Which AI agent protocols are actually used? MCP vs A2A vs x402

The adoption scoreboard

Most protocol comparisons list features. This one scores evidence of use, with dates. "Backed" and "used" are different claims — the gap between them is the whole story of 2026's agent stack.

Sources: Stripe newsroom (Sep 2025, Dec 11, 2025); Checkout.com industry coverage of the March 2026 checkout retreat; Artemis on-chain data on x402 as covered in protocol-landscape reporting (Crossmint, 2026); Linux Foundation announcements (Jun 2025); OpenAI and Google DeepMind MCP adoption announcements (Mar–Apr 2025).

What our own server logs add

This site's operators run three MCP servers listed on the official registry, across three properties. What we actually observe first-hand, as of August 2026: steady infrastructure interest — AI crawlers fetching llms.txt (about 180 fetches/day on one property) and spec-correct GET probes against MCP endpoints — but near-zero real agent tool-calls. One caution we learned the hard way and pass on: our first "agent adoption" signals turned out to be our own CI smoke tests. Anyone reporting agent-traffic growth should check whether the parameters repeat verbatim before believing it.

That first-hand picture matches the scoreboard: the plumbing (MCP, registries, llms.txt) is being adopted by platforms and crawlers faster than end-user agents are showing up to use it.

The enterprise reality check

The demand side explains the thin usage. MIT's Project NANDA found 95% of enterprise GenAI pilots deliver zero measurable P&L return (Fortune, Aug 18, 2025). Gartner projects over 40% of agentic-AI projects will be canceled by end-2027 and estimates that of thousands of vendors claiming "agentic," only ~130 have real agentic capability (Forbes, Jul 7, 2026). A protocol war fought above a market this immature produces exactly what the scoreboard shows: standards racing, usage crawling.

What would flip this verdict

• x402: sustained real (non-test) volume above ~$1M/day, or a first content-site case earning >$500/mo — we re-check the on-chain data around November 2026.
• Agentic checkout: an in-chat checkout shipping again at scale with published merchant volumes.
• A2A: verifiable cross-vendor production traces (not partner-logo announcements).
• MCP: the opposite risk — if registry growth stalls while probes stay bot-only, "real adoption" would need a downgrade to "infrastructure adoption."

Frequently asked questions

MCP (Model Context Protocol). It is the only agent protocol with cross-vendor production adoption you can verify: OpenAI adopted it in March 2025, Google DeepMind in April 2025, an official public server registry is live, and server operators can see agent and crawler probes in their own logs.

Barely. On-chain data (Artemis) showed transactions collapsing roughly 92% from about 731,000 to about 57,000 per day between December 2025 and February 2026, and by mid-2026 daily volume was on the order of $28,000 — with roughly half of it looking like test transactions. The protocol exists; a real market does not yet.

The first version retreated. OpenAI launched Instant Checkout with Stripe in September 2025 and relaunched Buy-it-in-ChatGPT in February 2026, but industry coverage reports it dropped in-ChatGPT checkout in March 2026. The payments standards war (ACP, AP2, TAP, x402) continues, but observable consumer usage remains thin.