Protocols Feature Matrix

Table of Contents

This matrix compares the five protocols profiled in this section, A2A, ACP, Agent Host Protocol, AGENTS.md, and MCP, so the whole interoperability stack can be read in one table. Everything below was re-verified against live sources on 2026-09-13.

The five do not compete, they stack (repo-to-agent, editor-to-agent, agent-to-tool, agent-to-agent, client-to-session), and adoption falls with every step up that stack, which is why I call AGENTS.md and MCP defaults, ACP a rising bet, AHP a bet underwritten by VS Code’s own distribution, and A2A an enterprise convention the coding-agent world can keep ignoring.

Legend: ✓ supported, ✗ not supported, ~ partial or conditional, ? not verified as of the date above. Each column links to the full research note; every cell below traces to a source cited there or in the references.

The matrix
#

Feature	A2A	ACP	Agent Host Protocol	AGENTS.md	MCP
Kind	wire protocol	wire protocol	wire protocol	file convention	wire protocol
Originated by	Google (2025-04)	Zed, with JetBrains	Microsoft (2026-03)	OpenAI-led (2025-08)	Anthropic (2024-11)
Steward	AAIF (2026-08), TSC	vendor-neutral org	Microsoft, no foundation	AAIF	AAIF
Spec license	Apache-2.0	Apache-2.0	MIT	MIT	MIT
Maturity	v1.0.1 (2026-05)	version 1, v2 draft, remote WIP	v0.9.0 (2026-08-28)	unversioned, de facto standard	dated revisions (2026-07-28)
What it connects	agent-to-agent	editor-to-agent	client-to-session	repo-to-agent	app-to-tools
Adoption in this section	✗ none native	~ growing (OpenCode, JetBrains, Zed)	~ VS Code reference host	~ most, Claude Code holdout	✓ near-universal
Transport or location	HTTP, gRPC, JSON-RPC	JSON-RPC over stdio	URI channels on a standalone sessions server	Markdown at repo root	JSON-RPC, stdio to sse
Official SDKs	✓ six	✓ five	✓ six	✗ none needed	✓ any language
Criticism recorded	redundant with MCP	sprawl, flattened UX	more sprawl, one vendor’s governance	weak efficacy evidence	tool poisoning, supply chain

Reading the matrix
#

Governance converged faster than adoption, and every protocol that went neutral did so only after it had already won or stalled. Google donated A2A to the Linux Foundation in June 2025, Anthropic and OpenAI donated MCP and AGENTS.md to the AAIF the same day (2025-12-09), and A2A itself joined the AAIF as a Growth Stage project on 2026-08-27, so three of the five now sit in the same foundation. ACP is stewarded by Zed and JetBrains under a vendor-neutral organization with no foundation home, AHP stays inside the microsoft org with repo-first governance, so the stack now has two non-foundation columns, and ACP remains the one compounding fastest in editors. I read this as governance following adoption, not causing it.

Adoption falls as the protocol climbs the stack, and the file convention beat every wire protocol to default status. MCP is table stakes across the harness and surface matrices; AGENTS.md counts more than 60,000 carrying projects with one glaring holdout (Claude Code); ACP rides OpenCode, JetBrains, Zed, and a Copilot CLI preview; A2A has no native speaker among this index’s harnesses, only community setups near Gemini CLI; AHP is nearly six months old with 328 stars as of 2026-09-13, but its reference host ships inside VS Code and the VS Code team has said publicly it is rebuilding its agent infrastructure on the protocol. The SDK download ratio recorded in the A2A note, 10.9M monthly versus 257M for MCP, is the gap in one number, and AHP’s roughly 300k combined downloads (about 219k crates plus 80k npm) show the same order-of-magnitude distance from the top.

The consolidations the notes record happened in opposite corners, and neither touched the other’s territory. IBM’s Agent Communication Protocol (the other ACP, the source of the name collision) merged into A2A in August 2025 under LF AI and Data. JetBrains folded its internal Junie protocol into the Agent Client Protocol instead. Both mergers cut the rival count in 2025 while leaving the enterprise-remote and local-editor layers separate.

Each recorded criticism is a different species of doubt, and the pattern favors the incumbents. MCP is criticized for risks it creates (tool poisoning, rug pulls, a 10,000-server supply chain); AGENTS.md for whether it helps at all (an ETH Zurich study found no success-rate gain and over 20% added inference cost, against Vercel’s counter-evidence); ACP for UX flattening and protocol sprawl; AHP for being a fifth protocol to track under one vendor’s governance; A2A for whether it should exist at all. The mature protocols get attacked for their risks, the young ones for their reason to exist.

Choosing from the matrix
#

Exposing tools or data to any agent: MCP, the default with a supply-chain asterisk.
Choosing harness and editor independently: ACP, the cheapest portability insurance in the stack.
Any repository an agent touches: commit a short AGENTS.md, commands and conventions first.
Delegating work across vendors or departments: A2A, provided both sides run enterprise platforms.
Attaching a second client to a live agent session: AHP, provided v0.9.0 churn and Microsoft’s stewardship are acceptable.
Wiring sub-agents inside one framework: none of these, native primitives or MCP are simpler.