Agent Protocols

Anthropic, developed in collaboration with HHMI Janelia Research Campus

A shared specification that lets AI agents safely operate physical devices such as robotic arms, microscopes, liquid handlers, and lasers in scientific and manufacturing settings. It defines a small set of read and write primitives (for example 'get temperature', 'set temperature') that any hardware device with a programmable interface can implement, and makes each device discoverable in a standard format so agents and devices can find and address each other without a bespoke translator program. It is model-agnostic and reachable by any agent harness over standard protocols such as MCP.

Algorand Foundation (spec editors: Bruno Martins, Michael Feher, Marc Vanlerberghe); reference wallet implementation by Pera Wallet, Lda (operated by Hipo, a company independent of the Algorand Foundation)

A peer-to-peer, authenticated messaging protocol that lets an AI Agent request a cryptographic signature from a human Controller without the Agent ever gaining access to the Controller's private keys. Messages are DIDComm v2.0-compliant and travel over a Liquid Auth (FIDO2/WebAuthn) established WebRTC DataChannel, with no central relay or blockchain required to operate. It ships with a reference OpenClaw plugin and native AC2 Wallet apps for Android and iOS.

Microsoft

A protocol that lets multiple clients share a synchronized view of a single AI agent session, so a coding agent's session can persist across separate editor windows, remote connections, or entirely different clients rather than being pinned to the process that started it. It runs a standalone, portable session server that clients subscribe to over URI-addressed channels (sessions, chats, terminals, changesets); the server maintains state via immutable state and pure reducers, stamps every mutation with a monotonic sequence number, and broadcasts the ordered stream so every connected client stays in sync. Its own documentation describes it as composing with, not replacing, the Agent Client Protocol: AHP coordinates multiple clients over a shared session, while ACP handles the point-to-point communication between one client and one agent.

Stripe and Tempo (Tempo Labs), joint authors

An open standard for machine-to-machine payments that repurposes HTTP 402 Payment Required, similar in mechanism to x402: a client requests a protected resource, the server responds 402 with a WWW-Authenticate: Payment challenge describing price and accepted methods, the client fulfills payment and retries with an Authorization: Payment credential, and the server validates the credential and grants access, all within a single request/retry cycle. It runs live on Tempo's payments-focused layer-1 blockchain and is payment-method agnostic, supporting stablecoins on Tempo, cards through Stripe, and custom payment methods, with each method built as a first-class protocol extension rather than routed through a separate facilitator service.

Mastercard, co-developed with Google, transitioning to the FIDO Alliance

An open specification for cryptographic agent authorization in commerce that creates tamper-evident delegation chains binding an AI agent's actions to the parameters a human actually approved. It links three parties into a signed delegation chain, an identity issuer, the intent-setting user, and the acting agent, using SD-JWT based credentials with constraint claims (amount bounds, merchant allowlists, budget caps, recurrence terms) so that any party downstream can cryptographically verify what an agent was authorized to do without re-deriving trust from scratch.

MCP Steering Group, jointly proposed by Anthropic and OpenAI with the MCP-UI open-source maintainers

An official extension to the core Model Context Protocol that lets MCP servers declare interactive HTML interfaces (charts, forms, dashboards, media viewers) that render inline inside MCP host conversations such as Claude or ChatGPT. A tool description points to a ui:// resource; the host fetches it and renders it in a sandboxed iframe, and the app talks back to the host over postMessage using its own ui/-prefixed JSON-RPC dialect that shares some methods with core MCP. It began as SEP-1865, proposed jointly by Anthropic and OpenAI together with the MCP-UI open-source maintainers; MCP-UI's own packages now serve as an implementation of the standardized spec rather than a separate standard.

Google, co-developed with retail and payments partners

An open standard for agentic commerce that spans the full shopping journey, discovery, cart building, checkout, order management and post-purchase support, so that AI agents, consumer surfaces (Search, Gemini, YouTube) and businesses can transact without bespoke integrations for every retailer. Businesses publish a JSON manifest of the 'capabilities' they support (such as Checkout or Identity Linking) at a well-known endpoint, letting agents discover features and endpoints dynamically. It is explicitly designed to interoperate with, rather than replace, Agent2Agent (A2A), the Agent Payments Protocol (AP2), and the Model Context Protocol (MCP).

Google, with contributions from CopilotKit and the open-source community

An open, transport-agnostic protocol and component format that lets agents describe rich, updatable user interfaces as a declarative JSON payload rather than executable code. A client-side catalog of trusted, pre-approved components (Card, Button, TextField, and similar) is the only vocabulary an agent may reference, so the agent requests what to render while the host's own native widgets (React, Flutter, SwiftUI, Angular, or web components) do the actual rendering. It ships with bindings for common transports including A2A, AG-UI, MCP, SSE with JSON-RPC, WebSockets, and REST rather than mandating one of its own.

OpenAI and Stripe (founding maintainers)

An open standard, co-developed by OpenAI and Stripe, for agent-initiated checkout: it defines how a buyer's AI agent collects payment and order details, hands a narrowly-scoped payment token to the merchant, and lets the merchant charge that token through any compliant payment service provider while the merchant keeps merchant-of-record status and control over pricing, fulfillment and branding. It comprises three specs: an Agentic Checkout Spec for merchants, a Delegated Payment Spec for payment service providers, and a Product Feed Spec for product data.

Google, transitioning to the FIDO Alliance

An open protocol developed with leading payments and technology companies to securely initiate and transact agent-led payments across platforms. AP2 does not move money itself; it establishes a payment-method-agnostic trust layer built on cryptographically signed 'mandates' (Intent Mandate, Cart Mandate, Payment Mandate) that create a verifiable, non-repudiable record of what a user actually authorized an agent to buy, so merchants and payment providers can trust an agent-initiated transaction the same way they trust a human-initiated one.

Zed Industries and JetBrains, jointly, under an interim governance model stated to be moving toward an independent foundation

A protocol that standardises communication between code editors (or any client UI) and AI coding agents, decoupling agents from the surfaces that host them. Agents run as subprocesses and exchange JSON-RPC messages with the editor for sessions, streaming updates, permission requests, file operations, and terminal access. Analogous to what LSP did for language tooling.

UTCP Contributors, a community project with core maintainers from Bevel Software

An open specification that lets AI agents discover and call tools directly over the tool's own native protocol (HTTP, CLI, WebSocket, gRPC, text, and others) instead of routing every call through a dedicated proxy server. A UTCP 'manual' is a JSON manifest describing available tools and how to reach them; the agent reads the manual once and then talks to the underlying API, CLI, or service directly, reusing whatever auth and infrastructure that endpoint already has.

Microsoft (originated by R.V. Guha's team; the open-source reference implementation now lives under the community nlweb-ai GitHub organization)

An open protocol that gives a website a natural-language query interface for both human visitors and AI agents. An NLWeb-enabled site exposes an /ask endpoint that accepts a plain-language question and returns structured JSON answers built on the Schema.org vocabulary the site likely already publishes, rather than requiring agents to scrape and parse HTML pages designed for humans. Every NLWeb instance also runs as a Model Context Protocol server so MCP-speaking agents can query the same content.

CopilotKit

An open, lightweight, event-based protocol that standardizes how AI agents connect to user-facing applications, so any agent backend (LangGraph, CrewAI, Mastra, Google ADK, and others) can stream messages, tool calls, state patches, and lifecycle signals into any frontend without a bespoke integration. Created by CopilotKit, it sits at the agent-to-user-interface layer and is designed to complement tool-access protocols like MCP rather than replace them.

x402 Foundation (Linux Foundation), originally created by Coinbase

An open, internet-native payment standard that repurposes the long-dormant HTTP 402 'Payment Required' status code so that clients (including AI agents) and servers can exchange machine-to-machine payments directly inside the normal HTTP request/response cycle, with no accounts, sessions, or out-of-band checkout flow required. It defines standardized payment schemes (starting with an 'exact amount' scheme) that can settle over multiple networks, both crypto (stablecoins on Base, Solana and others) and, as the standard matures, traditional card and bank rails.

Cloudflare, with co-authors from Google and Amazon; as of September 2026 the core signing draft is a working-group document of the IETF's newly chartered webbotauth working group, while the companion registry draft remains an individual submission

A profile of RFC 9421 (HTTP Message Signatures) that lets automated HTTP clients, including AI agents, cryptographically sign outbound requests so that servers can verify the caller's identity instead of relying on spoofable signals like User-Agent strings or IP allowlists. A companion draft defines a 'Signature Agent Card', a JSON metadata document a signing agent publishes describing its identity, purpose, rate expectations, and public keys, drawn from the OAuth Dynamic Client Registration Metadata namespace.

Linux Foundation (A2A Project, accepted 2026-08-27 as a Growth Stage project of the Linux Foundation's Agentic AI Foundation), originally developed by Google

An open, vendor-neutral standard that lets AI agents built on different frameworks discover each other's capabilities, exchange tasks, and collaborate on multi-step work without sharing internal memory, tools, or proprietary logic. Agents publish an Agent Card, a JSON metadata document describing identity, skills, and service endpoints, and exchange stateful Tasks over JSON-RPC, gRPC, or HTTP/REST bindings, with Server-Sent Events for streaming updates.

Linux Foundation (LF AI & Data), originally IBM Research / BeeAI

IBM Research's protocol for agent-to-agent messaging, built to power the open-source BeeAI platform so agents from different frameworks and languages could discover each other, exchange multimodal messages (text, code, files, media), and coordinate through a REST-based API with agent manifests advertising capabilities. In August 2025 the ACP team wound down independent development and contributed its technology and expertise into the Agent2Agent Protocol under the Linux Foundation.

Model Context Protocol, a Series of LF Projects, LLC (part of the Linux Foundation's Agentic AI Foundation); Anthropic remains an active maintainer

An open protocol that standardises how AI applications connect to external tools, data sources, and prompts. Servers expose tools, resources, and prompt templates over a JSON-RPC interface; clients (chat apps, IDEs, agents) consume them without bespoke integrations. Often described as the USB-C port for AI applications.

Independent open-source community, maintained by GaoWei Chang

A community-maintained, three-layer protocol suite (open internet infrastructure, an identity and encrypted-communication layer, and an application protocol layer) that aims to give autonomous agents verifiable decentralized identity, discovery, and negotiation without a central authority. It builds on W3C standards such as Decentralized Identifiers via a did:wba method, and proposes a meta-protocol layer that would let agents negotiate bespoke communication protocols on the fly, though that meta-protocol layer remains an unreleased draft.

Answer.AI (proposed by Jeremy Howard)

A proposed standard markdown file, served at /llms.txt on a website's root (or a subpath), that gives AI agents and LLMs a curated, machine-readable summary of a site's content and links to further markdown versions of its pages. It targets the same problem robots.txt and sitemap.xml solve for crawlers, but for language models with limited context windows that struggle to parse HTML built for human navigation and ads.