Tech-watch note by Didier Girard dated August 2, 2026, prompted by a colleague's question ("what is ACP?") that turns out to address a problem that is not terminological but documentary. Three protocols compete for the acronym, with no technical overlap whatsoever: Agent Client Protocol (client ↔ agent — Zed, August 2025, JSON-RPC 2.0 over stdio, Apache-2.0, "what LSP did for languages"), Agentic Commerce Protocol (agent ↔ merchant — OpenAI + Stripe, Sept. 29, 2025, facing Google's UCP of Jan. 11, 2026 backed by AP2), and Agent Communication Protocol (agent ↔ agent — IBM Research / BeeAI, marginal but polluting searches). The core of the note is not the disentangling but its observed failure: the author searches "ACP" in his tech-watch knowledge base and gets twelve results, all about the commerce protocol, zero about Zed's — "our watch agents had indexed the acronym without disambiguating it".
By **Didier Girard** — auteur de la note. Écrit ici depuis la position de **praticien de la veille outillée** : le déclencheur est une question de collègue// Source internal document /Reading 2 min/.md// Auto-verified translation
Tech-watch note dated August 2, 2026, born from a colleague's question — "what is ACP?" — which the author shows has no simple answer: three protocols compete for the acronym, with not the slightest technical overlap.
Agent Client Protocol connects a client to an agent. Introduced by Zed in August 2025, it does for agents what LSP did for languages: it decouples the editor from the agent. Before, N editors × M agents required N×M custom-built integrations; afterward, each side speaks the protocol and N+M suffices. JSON-RPC 2.0 over stdio, Apache-2.0. The note points out that the protocol has left its creator's orbit — its own organization, an agent registry, a versioned specification, a JetBrains implementation.
Agentic Commerce Protocol has nothing to do with it: it connects an agent to a merchant (discovery, cart, payment). Announced by OpenAI and Stripe on September 29, 2025, it faces Google's UCP (January 11, 2026), backed by AP2 for payment. The stake: the "Visa/Mastercard" layer of agentic commerce. Agent Communication Protocol (IBM Research / BeeAI), agent-to-agent, rounds out the picture and pollutes searches.
The real subject is documentary. The author searches "ACP" in his tech-watch database: twelve results, all about the commerce protocol, zero about Zed's. The indexing agents had processed the acronym without disambiguating it. Hence the rule adopted: "never index a bare acronym" — the entity is the full name, the acronym is only an alias, here carried by three distinct entities. In passing, the note clears up a neighboring confusion: MCP connects an agent to its tools, ACP connects a client to an agent, and the two stack.
The concrete case is Buzz, published by Block on July 21, 2026 under Apache-2.0: a self-hostable workspace on Nostr where humans and agents share the same channels, each participant being a keypair and every event — message, workflow step, git push — being signed in an append-only log. The agent architecture is entirely protocol-based (buzz-acp, buzz-agent, buzz-dev-mcp), hence the agnosticism: Goose, Claude Code, and Codex through the same harness, and Hermes connected without a single line of code on Block's side. "N+M instead of N×M, running in production."
The punch line concerns the Claude subscription versus third-party agents, after a turbulent 2026 (OAuth block, separate credits announced then suspended the day they took effect). The rule adopted separates ordinary, individual use from routing other people's requests. Its formulation applies beyond this case: "the distinction is not legal, it is architectural: who is consuming, and on whose behalf" — best settled at design time rather than by reading the terms of service.
Key takeaways
The disentangling, worth memorizing once and for all. | Acronym | Full name | Connects | Origin | |---|---|---|---| | ACP #1 | Agent Client Protocol | a client to an agent | Zed, August 2025 | | ACP #2 | Agentic Commerce Protocol | an agent to a merchant | OpenAI + Stripe, Sept. 29, 2025 | | ACP #3 | Agent Communication Protocol | an agent to an agent | IBM Research / BeeAI | → No technical overlap. And the surface resemblance is maximal ("protocol", "agents", "open source", "2025-2026"), which is precisely what makes it a trap.
The analogy that lands. ACP #1 is to the agent what LSP is to the language. Before, N editors × M agents = N×M custom-built integrations; afterward, each side speaks the protocol and N+M suffices. It's the classic interoperability argument, and it's the only one that matters for deciding whether to adopt it.
MCP vs ACP — the table to reuse as-is.MCP connects an agent to its tools and data; ACP connects a client to an agent. They stack: the client speaks ACP to the agent, the agent speaks MCP to its tools. A frequent confusion, cleared up in two lines.
⭐ The curation rule — the note's real deliverable."never index a bare acronym". The canonical entity is the full name; the acronym is only an alias, and an alias can be carried by several entities. This is directly applicable to this repository — see [docs/reference/ontologie-kg.md](docs/reference/ontologie-kg.md) (disambiguation rules) and scripts/entity_aliases.tsv.
⭐⭐ VERIFICATION CARRIED OUT ON THIS CORPUS (2026-08-02) — the thesis holds, and worse. 1. "Zero about Zed's protocol" → confirmed exactly. At the time of verification, the string "Agent Client Protocol" was absent from fiches/ and kb/; "Zed" appeared only in an unrelated fiche ([[powered-by-claude-anthropic-partners-2025-07-09]]). → Gap closed the same day by [[agentclientprotocol-introduction-2026-08-02]] (primary source), [[dethlefsen-zed-anthropic-subscription-changes-2026-05-14]], and [[sawers-thenewstack-anthropic-pause-agent-sdk-subscription-2026-06-16]]. 2. "Twelve results" → not reproducible by grep: 3 fiches contain the token ACP ([[thilen-opascope-ai-shopping-assistant-agentic-commerce-protocols-2026-02-10]], [[marette-agentic-commerce-optimization-acp-ucp-2026-02-23]], [[ragsdale-merit-open-agentic-commerce-protocols-2026-03-19]]), for 32 occurrences. The "twelve" likely comes from a semantic search engine, not a grep. The number doesn't check out; the qualitative conclusion does — and that's the one that matters. 3. The trap was already loaded, not merely possible: contrary to the proposed rule, the bare acronym ACP was already typed as an entity in the graph, subject of triples in two fiches ([[thilen-opascope-ai-shopping-assistant-agentic-commerce-protocols-2026-02-10]], [[ragsdale-merit-open-agentic-commerce-protocols-2026-03-19]]), alongside a variant Agentic Commerce Protocol (ACP) — i.e. three entities for a single protocol. Had a fiche on Zed's ACP typed ACP as an entity, the two protocols would have merged into a single KB page. → Fixed on August 2, 2026: the variants were normalized to Agentic Commerce Protocol, and symmetrically UCP → Universal Commerce Protocol and ACS → Agentic Commerce Suite. The note's rule was therefore not preventive here: it was corrective. 4. ⚠️ A defect already materialized, and a more serious one: the page kb/Agentic-Commerce-Protocol.md carries the attribute "category: Google agentic commerce protocol" and the relation Google created → Agentic Commerce Protocol (confidence 0.99). Yet the ACP is OpenAI + Stripe; Google's protocol is UCP (+ AP2). The confusion the note describes between two ACPs has already produced, in this corpus, an attribution error between OpenAI's ACP and Google's protocol. → Fix to be applied to the source fiches (since the KB is generated, it cannot be corrected by hand).
Governance of ACP #1 — a nuance to add. the note writes "Zed launched the thing, then let it go". The opening move is real and verifiable (its own organization, versioned specification, public registry, JetBrains implementation). But public sources describe more of a shared Zed × JetBrains governance — the agent registry was co-launched by the two in January 2026, and ACP is still presented on Zed's side as its open standard and the flagship feature of Zed 1.0. Openness, yes; dispossession, no. A nuance with no consequence for the interoperability argument, but not to be overplayed when speaking about it. ⚠️ Detail since disproven, by the primary source: the version number ACP 1.2 is not corroborated — the specification's navigation shows v1 (Latest) and v2 (Draft), cf. [[agentclientprotocol-introduction-2026-08-02]]. Cite v1 / v2 draft. The registry URL cdn.agentclientprotocol.com/registry remains unconfirmed (public sources point to agentclientprotocol.com/registry).
Buzz — the concrete case, verified and accurate. published by Block on July 21, 2026, Apache-2.0, repository github.com/block/buzz. A self-hostable workspace on Nostr; every human or agent participant is a keypair (Schnorr); every message, workflow step, and git push is a signed event in an append-only log. Three protocol building blocks: buzz-acp (a harness translating Buzz events to an agent, in ACP over stdio), buzz-agent (an ACP agent calling an LLM and using MCP tools), buzz-dev-mcp (an MCP server for shell + file editing). Three harnesses shipped out of the box: Goose (Block), Claude Code (Anthropic), Codex (OpenAI). A self-hostable relay — an explicit data sovereignty argument at Block.
⭐ Why Buzz is the right illustration. it's not "a product that uses ACP", it's the demonstration that the decoupling holds. Three agents from three competing vendors plugged in through the same harness, and a fourth (Hermes, Nous Research) connected without Block writing a single line. The N+M promise stops being an architecture argument and becomes an observable fact. (Nuance: on Hermes's side, the public source frames the addition as an ACP server mode opening up Zed, JetBrains, Neovim, and others — the benefit for Buzz is a consequence of ACP, not a targeted integration. Which reinforces the argument rather than weakening it.)
The enterprise-architecture reading. Buzz provides a sovereign, auditable multi-agent substrate, with no third-party SaaS dependency — signed log, self-hosted relay, cryptographic identities. It's one possible answer to the question of "where human-agent collaboration will live in the information system". Worth relating to the convergence of hyperscaler agent platforms analyzed in [[janakiram-agent-platform-portability-contract-2026-07-20]]: Buzz is the self-hosted option for the same portability need.
Claude subscription and third-party agents — 2026 timeline as reported. (⚠️ taken from the note's sources, not re-verified here) : Jan. 9 silent technical block on subscription OAuth tokens outside official tools → Feb. 17-20 formalization in the docs and the Consumer Terms (the OpenClaw episode) → Apr. 4 full enforcement → May 13-14 announcement of separate Agent SDK credits ($20 Pro / $100 Max 5x / $200 Max 20x) effective June 15 → June 15: the change suspended the very day it took effect. State reported as of August 2, 2026: the Agent SDK, claude -p, and third-party applications built on it draw from the subscription's limits; the separate credit does not exist. → This state is time-bound by construction: re-verify it before any decision.
⭐⭐ The rule that will outlast the terms of service — the note's best idea. the line drawn by Anthropic separates "ordinary, individual" use from routing other people's requests via Free/Pro/Max credentials. Translation: an owner-only agent consumes your subscription for you — within bounds; an anyone agent in a shared channel routes your colleagues' requests through your account — out of bounds, and that's the moment to switch to a usage-billed API key. The distinction is not legal, it is architectural: who is consuming, and on whose behalf. To be settled at design time, not by reading the terms of service. ⭐ A connection worth making : this is exactly the ambient authority from [[valente-zalewski-beyond-zero-enterprise-security-ai-era-2026-07-20]], transposed from authorization to billing. An agent that inherits its human's permissions and acts on behalf of a third party poses the same boundary problem, whether it's accessing a document or consuming a quota. The question "on whose behalf?" is the same invariant in both cases.
Meta / to link. on the commerce side, the three fiches carrying the homonymy — [[ragsdale-merit-open-agentic-commerce-protocols-2026-03-19]], [[marette-agentic-commerce-optimization-acp-ucp-2026-02-23]], [[thilen-opascope-ai-shopping-assistant-agentic-commerce-protocols-2026-02-10]] — plus [[nrf-2026-commerce-agentique-ucp-deep-research-2026-01-13]] and [[google-agentic-commerce-ap2-payment-protocol-2025-09-16]] on the Google side. On the Block/Goose side: [[block-goose-mcp-ui-future-agentic-interfaces-2025-08-25]] and [[mcp-ui-future-agentic-interfaces-goose-2025-08-25]]. On the open standardization of the agent layer: [[openai-agentic-ai-foundation-linux-2025-12-09]]. ⚠️ Graph guideline arising from this fiche : never create an ACP entity. The three canonical entities are Agent Client Protocol, Agentic Commerce Protocol, and Agent Communication Protocol; "ACP" must only ever appear as an alias explicitly attached to one of them.
Attributed claims
three distinct protocols share the acronym ACP with no technical overlap
— Didier Girard
the boundary between individual use and routing others' requests is architectural, not legal — who consumes, and on whose behalf
— Didier Girard
Pro and Max limits assume ordinary, individual usage of Claude Code and the Agent SDK
— Anthropic
The knowledge graph extracted from this fiche — 11 entities, 27 relations.