Agent Protocols

Transport layer

JSON-RPC 2.0 over stdio and streamable HTTP, redesigned as a stateless request/response protocol in the 2026-07-28 revision: the initialize handshake is removed, every request self-describes its protocol version, and standard headers (Mcp-Method, Mcp-Name) let gateways route without parsing bodies. HTTP+SSE is now formally deprecated (not just discouraged).

Key features

• Tools: model-invocable functions with typed schemas and structured output
• Resources: application-readable context (files, data) addressed by URI
• Prompts: reusable prompt templates exposed by servers
• Multi Round-Trip Requests (MRTR): servers request user input mid-call via an InputRequiredResult, replacing the deprecated server-initiated Sampling and Elicitation calls
• OAuth 2.1 based authorization for remote servers, now favoring Client ID Metadata Documents over Dynamic Client Registration
• Optional extensions framework (Tasks, MCP Apps) layered on top of the stateless core

Timeline

1. 2024-11-25Anthropic open-sources MCP with initial spec, SDKs, and reference servers
2. 2025-03-26Spec revision: streamable HTTP transport replaces HTTP+SSE, OAuth 2.1 authorization framework, tool annotations
3. 2025-03-27OpenAI announces MCP adoption across Agents SDK, ChatGPT desktop, and Responses API
4. 2025-06-18Spec revision: elicitation, structured tool output, resource-server authorization clarifications, protocol version headers
5. 2025-09-08Official MCP Registry launches in preview as the canonical index of public servers
6. 2025-11-25Spec revision adds experimental Tasks, OpenID Connect discovery, and incremental OAuth scope consent, and formalizes MCP's governance structure
7. 2025-12-09Anthropic donates MCP to the Linux Foundation's new Agentic AI Foundation, becoming Model Context Protocol, a Series of LF Projects, LLC
8. 2026-01-26MCP Apps ships as the first official MCP extension for interactive UI, co-authored by Anthropic and OpenAI with the MCP-UI maintainers
9. 2026-07-28Spec revision makes MCP a stateless request/response protocol (MRTR pattern, header-based routing, cacheable list results); Tasks moves into an official extension and Roots, Sampling, and Logging are deprecated

Roadmap

The 2026 roadmap (blog.modelcontextprotocol.io/posts/2026-mcp-roadmap) sets four priorities: transport evolution and scalability so servers can scale horizontally without holding state; agent communication (the specific scope was not independently verified beyond the priority label itself); governance maturation (a documented contributor ladder and delegation of SEP review to Working Groups); and enterprise readiness (audit trails, SSO-integrated auth, gateway behavior, configuration portability). No GA date is stated for the registry, which remains in preview.

Adoption signals

Adopted far beyond Anthropic: OpenAI announced MCP support in March 2025, followed by Google DeepMind and Microsoft. Thousands of community servers; an official MCP Registry launched in preview in September 2025. Support in Claude, ChatGPT, Cursor, VS Code, Windsurf, and most major agent products.

The de facto standard for agent-to-tool connectivity as of 2026. As of December 2025 MCP is no longer solely Anthropic-governed: it was donated to the Linux Foundation's Agentic AI Foundation as Model Context Protocol, a Series of LF Projects, LLC, alongside Block's goose and OpenAI's AGENTS.md. Anthropic remains an active maintainer and infrastructure contributor. The MCP Registry (preview since September 2025) is still in preview as of August 2026, with no GA date announced.

Transport layer

Protocol-agnostic JSON manual (call templates) pointing at the tool's native transport: HTTP, CLI, WebSocket, text/SSE, gRPC, or an MCP server; no wrapper/proxy server is required to relay the call

Key features

• JSON 'manual' format for describing tools and their native call templates
• Direct calls to a tool's own endpoint, avoiding a proxying middle-hop server
• Pluggable transport support: HTTP, CLI, WebSocket, text, gRPC, and MCP
• Reuses the tool's existing authentication (API key, OAuth, custom) rather than reimplementing it
• Tool discovery mechanism separate from the call mechanism
• Official SDKs in Python, TypeScript/JS, and a growing set of framework adapters (LangChain, Pydantic AI)

Timeline

1. 2025-07-02utcp-specification repository created on GitHub
2. 2025-07-17UTCP publicly launched on Product Hunt as 'the scalable, secure alternative to MCP for tool calling'

Adoption signals

Launched publicly on Product Hunt on July 17, 2025 as an alternative to MCP. The python-utcp reference implementation has a few hundred stars on GitHub. Third-party projects have built on it, including Bevel Software's Hexis, which uses versioned UTCP manuals as git-backed tool integrations. Adoption is early-stage community traction rather than large-vendor endorsement.

Competes against

Model Context Protocol

Positions itself explicitly against MCP's client/server proxy model, arguing it removes a 'double hop' and a wrapper-server requirement. Governance is community-run rather than a single vendor, though Bevel Software staff hold admin and distribution-maintainer roles.

Transport layer

Reuses existing web infrastructure (HTTP, DNS, TLS, CDN) as the base transport, layered with did:wba-based identity and encrypted messaging

Key features

• did:wba and WNS: verifiable agent identity, readable handles, and cross-domain trust roots
• Agent description and discovery protocols for publishing and finding agent capabilities
• Meta-protocol for dynamic negotiation of communication protocols between agents (currently draft, unreleased)
• Payment protocol layer for agent-to-agent value exchange
• Three-layer architecture separating internet infrastructure, identity/communication, and application protocols

Timeline

1. 2024-10-24v0.0.1 initial release: first published version of the ANP specification and repository
2. 2025-05-19v1.0 release: adds human-in-the-loop authorization, multiple DID support, and a peer-to-peer agent transaction framework
3. 2026-06-27v1.1 release: improved identity-authentication design, adds WNS naming support and Agent Instant Messaging specifications

Roadmap

The project's GitHub README lists an ANP 1.2 draft in progress covering multi-device messaging, and states the meta-protocol specification for dynamic negotiation between agents 'remains a draft and is not released yet,' identifying it as the steward's next major spec milestone.

Adoption signals

Open-sourced under the MIT license with an active GitHub community (a few thousand stars and a growing number of forks at time of review) and a companion AgentConnect SDK implementing did:wba authentication and protocol negotiation. Documented and compared against MCP, Agora, and LMOS in a July 2025 arXiv technical white paper, but adoption is largely within the Chinese open-source AI community rather than backed by a company or foundation.

Competes against

Agent2Agent Protocol

Emphasizes decentralized W3C DID-based identity as its core primitive rather than a central registry (A2A's Agent Card model) or Eclipse's centralized/decentralized registry split (lmos-protocol). Community-governed (an individual maintainer), unlike the foundation- or corporate-led entries in this category.

Transport layer

JSON-RPC 2.0, gRPC, and HTTP+JSON/REST protocol bindings, with Server-Sent Events for streaming

Key features

• Agent Cards: JSON documents (typically hosted at /.well-known/agent-card.json) advertising identity, skills, and auth requirements
• Task-based delegation: work items with a unique ID progress through a defined stateful lifecycle
• Multiple protocol bindings (JSON-RPC, gRPC, REST) so agents can interoperate regardless of transport preference
• Streaming updates and push notifications for long-running or background tasks
• Formal extension mechanism with a tiered promotion process for new capabilities
• Six official SDKs (Python, JavaScript, Java, C#/.NET, Go, Rust)

Timeline

1. 2025-04-09Google announces the Agent2Agent Protocol at Google Cloud Next, with backing from 50+ partners
2. 2025-06-23Google donates A2A to the Linux Foundation at Open Source Summit North America, establishing vendor-neutral governance
3. 2025-08-29IBM's Agent Communication Protocol (ACP) merges into A2A under the Linux Foundation umbrella
4. 2026-03-12A2A Protocol v1.0.0 released: first stable, production-ready version with breaking changes to the interaction protocol
5. 2026-05-28A2A Protocol v1.0.1 released: bug fixes for HTTP binding preferences and TaskStatus values

Roadmap

Steward roadmap (a2a-protocol.org/latest/roadmap) states priorities of expanding A2A extensions with SDK backing, fostering community-led development with standardized contribution processes, strengthening validation via the A2A Inspector and a Protocol Technology Compatibility Kit, and maintaining SDK support across six languages as adoption accelerates.

Adoption signals

Donated by Google to the Linux Foundation in June 2025 with a Technical Steering Committee spanning AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP, and ServiceNow. Reported to have a large and growing base of supporting organizations within its first year, and IBM's competing Agent Communication Protocol wound down independent development and merged its technology into A2A in August 2025.

Competes against

Agent Network Protocol

Peer-to-peer agent messaging and task delegation, distinct from model-context-protocol (tool access) and agent-client-protocol (editor-to-agent). Its predecessor-by-merger, IBM's Agent Communication Protocol, is tracked separately in this dataset as agent-communication-protocol with status superseded.

Transport layer

REST over HTTP with an OpenAPI specification; supported streaming and background/async responses

Key features

• Multimodal messaging: text, code, files, and media exchanged between agents
• Agent manifests for capability advertisement and discovery
• Session management for maintaining state across multi-turn interactions
• Interactive exchanges where agents can pause and request additional information
• Real-time, streaming, and background/async response modes
• Client libraries for Python and TypeScript, plus a BeeAI reference implementation

Timeline

1. 2025-04-09IBM contributes the BeeAI platform, and with it ACP, to the Linux Foundation
2. 2025-08-27i-am-bee/acp GitHub repository archived with a banner: 'ACP is now part of A2A under the Linux Foundation'
3. 2025-08-29LF AI & Data publishes 'ACP Joins Forces with A2A,' confirming ACP winds down independent development and contributes its technology to A2A

Adoption signals

Used as the interop layer for IBM's BeeAI platform; donated to the Linux Foundation alongside BeeAI in April 2025. After the June 2025 arrival of Google's A2A, the two projects' teams found overlapping goals and merged, with IBM's Kate Blair joining the A2A Technical Steering Committee. The GitHub repo was archived on August 27, 2025 with a migration guide pointing users to A2A.

Competes against

Agent2Agent Protocol

Not to be confused with Zed's Agent Client Protocol (agent-to-ui, unrelated collision on the ACP initialism) or OpenAI/Stripe's Agentic Commerce Protocol (payments, also unrelated). This entry uses the full unabbreviated id specifically to avoid that three-way ACP collision.

Transport layer

JSON-RPC 2.0 over stdio between client and agent subprocess

Key features

• Session management: create, load, and resume agent sessions from any client
• Streaming updates: token-level agent output and tool-call progress into the editor UI
• Permission requests: agents ask the client before running tools or shell commands
• File system proxying: agents read and edit files through the client, respecting unsaved buffers
• Terminal integration: agents can request client-managed terminal execution
• MCP-friendly: reuses MCP-style content blocks so agents can bridge both protocols

Timeline

1. 2025-08-27Zed announces ACP with Gemini CLI as the flagship external agent integration
2. 2025-09-15Community adapters bring Claude Code and other agents to ACP-capable editors
3. 2026-01-26ACP Agent Registry goes live in JetBrains IDEs for one-click agent installation
4. 2026-02-18Sergey Ignatov (JetBrains) becomes a Lead Maintainer of ACP alongside Zed's Ben Brandt, formalizing joint stewardship
5. 2026-03-09ACP Registry reaches Completed status, giving clients a standard way to discover, install, and configure compatible agents
6. 2026-07-20ACP v2 draft published, introducing breaking changes for richer session-state expression while v1 support continues alongside it

Roadmap

The 1.0 schema has already shipped and stabilized; the stated direction is now a v2 draft (published 2026-07-20) focused on expressing more session states, consolidating patterns found useful in v1, uniform message patching with stable IDs, and forward compatibility by default. Maintainers stress v2 is not yet stable, recommend gating it behind version negotiation and feature flags, and say v1 support should not be dropped when v2 support is added.

Adoption signals

Ecosystem has grown far beyond the original Zed and Gemini CLI launch: agentclientprotocol.com/overview/agents lists roughly 50 agents implementing ACP (Claude Agent, Cursor, Codex CLI, Gemini CLI, Goose, OpenHands, Qwen Code, and others), and agentclientprotocol.com/overview/clients lists about 13 editors and clients (Zed, JetBrains IDEs, Visual Studio, VS Code, Neovim, Emacs, Obsidian, and others). JetBrains partnered with Zed on ACP in October 2025 and became a joint steward; a shared ACP Agent Registry went live in JetBrains IDEs in January 2026 and reached Completed status in March 2026.

Not to be confused with IBM BeeAI's Agent Communication Protocol, which shares the ACP abbreviation but targets agent-to-agent messaging. The abbreviation collision is tracked in the topic validation checks.

Transport layer

Streams a single sequence of JSON events over standard HTTP (Server-Sent Events), with an optional binary channel for performance-critical cases; builds on HTTP and WebSockets as its base transports

Key features

• Streaming chat: live token and event streaming for multi-turn sessions
• Generative UI: agents can render both static and declarative interface components
• Shared state: event-sourced diffs and conflict resolution between agent and frontend state
• Human-in-the-loop: pause, approve, edit, retry, or escalate mid-flow
• Tool integration: tools can execute on either the frontend or backend
• Multimodality: typed attachments and real-time media (files, images, audio)

Timeline

1. 2025-05-13CopilotKit announces AG-UI with day-0 integrations for LangGraph, CrewAI, Mastra, and AG2
2. 2025-10-30CopilotKit reports AG-UI crossed 9,000 GitHub stars and that Google ADK joined the AG-UI ecosystem, alongside adoption from LangChain, AWS, and Microsoft

Adoption signals

Launched with day-0 integrations for LangGraph, CrewAI, Mastra, and AG2. First-party support has since grown to include Microsoft Agent Framework, Google ADK, AWS Strands Agents, Pydantic AI, Agno, and LlamaIndex, with OpenAI Agent SDK and Cloudflare Agents listed as in progress. The MIT-licensed GitHub repo ships SDKs in TypeScript, Python, Kotlin, Go, Dart, Java, Rust, Ruby, C++, and .NET.

Competes against

A2UIMCP Apps

Distinct from MCP (agent-to-tool) and from Agent Client Protocol (editor-to-coding-agent, JSON-RPC over stdio subprocess). AG-UI targets the general agent-to-frontend layer for any application, not specifically code editors. currentVersion (0.1.20) is the SDK package version on PyPI; the protocol itself has not published a separate spec version number as of this research.

Transport layer

postMessage between the host page and a sandboxed iframe, carrying a JSON-RPC dialect that shares some methods with core MCP (e.g. tools/call) and adds ui/-prefixed methods (e.g. ui/initialize)

Key features

• UI resource declaration: tool descriptions reference a ui:// resource the host can preload before the tool is even called
• Sandboxed iframe rendering: apps cannot access the parent page, cookies, or local storage
• Bidirectional communication: apps can call MCP tools and receive fresh data pushed from the host
• Capability negotiation: hosts control which tools, permissions (camera, microphone), and origins (via CSP) an app may use
• Framework agnostic: starter templates exist for React, Vue, Svelte, Preact, Solid, and vanilla JavaScript
• Delegated actions: apps can ask the host to route an outcome through the user's already-connected tools instead of building their own integrations

Timeline

1. 2025-11-21Anthropic and OpenAI publish the joint MCP Apps proposal (SEP-1865), developed with the MCP-UI maintainers
2. 2026-01-26MCP Apps ships as the first official MCP extension, live in Claude, Claude Desktop, VS Code Copilot, Microsoft 365 Copilot, Goose, Postman, MCPJam, and Archestra.AI, with ChatGPT support the same week

Adoption signals

Supported at launch by Claude, Claude Desktop, VS Code GitHub Copilot, Microsoft 365 Copilot, Goose, Postman, MCPJam, and Archestra.AI, with ChatGPT support following the same week. The community @mcp-ui/client React package and the MCP-UI project (mcpui.dev) provide a widely used implementation path alongside the official AppBridge SDK.

Competes against

Agent-User Interaction ProtocolA2UI

An extension of, not a replacement for, model-context-protocol (its parent entry in this dataset): it has its own dated spec version (2026-01-26), own repository, own transport (postMessage over iframe, versus MCP's stdio/streamable-HTTP), and own SEP, so it is tracked as a distinct entity even though model-context-protocol's own keyFeatures/timeline also mention it as an extension shipped under MCP's umbrella. MCP-UI (mcpui.dev) was the pre-existing community project whose design became the basis for this standardized spec.

Transport layer

Transport-agnostic JSON message format; documented bindings include A2A, AG-UI, MCP, SSE with JSON-RPC, WebSockets, and REST

Key features

• Security-first, non-executable format: agents can only reference components already registered in the client's trusted catalog
• Adjacency-list component model: components are a flat list linked by ID references, enabling progressive and asynchronous rendering
• Scoped data binding: absolute and relative data-model paths support template-based list rendering
• Two-way input binding: interactive components maintain a live read/write contract with the data model
• Client-side function registry: validation and formatting logic runs on the client without shipping executable code
• Framework-agnostic renderers: portable across web, Flutter, and native mobile clients

Timeline

1. 2025-12-15Google publishes A2UI as an open-source project for agent-driven, cross-platform generative UI
2. 2026-04-17Google publishes the A2UI v0.9 release, shifting to a prompt-first design embedded in the LLM prompt and adding sendDataModel plus a portable checks-based validation model

Roadmap

The a2ui.org specification pages describe v1.0 as a release candidate adding bidirectional function calls, single-message UI instantiation, and decoupled branding, while v0.9.1 remains the recommended production version until v1.0 stabilizes.

Adoption signals

Google positions A2UI alongside its own Opal and Gemini Enterprise products and the Flutter GenUI SDK; CopilotKit has published integration guidance combining A2UI with AG-UI. Apache 2.0 licensed and under active spec iteration (v0.8 legacy, v0.9.1 current, v1.0 release candidate as of mid-2026).

Competes against

Agent-User Interaction ProtocolMCP Apps

Widely glossed by press coverage as Agent-to-User Interface, though Google's own posts do not spell out the acronym expansion, so abbreviation is left null and the expansion is left unstated. Marked status draft per the topic's own pre-1.0 definition despite real production adoption, since Google itself labels v1.0 a release candidate, not final.

Transport layer

Positioned as an extension of the Agent2Agent (A2A) protocol and Model Context Protocol (MCP) rather than a new wire transport; payment authorization is carried as signed Verifiable Digital Credentials (mandates) exchanged between agents, merchants and payment providers

Key features

• Intent Mandate: captures a user's purchase instructions and constraints (price limits, timing) as an auditable, signed record
• Cart Mandate: locks in the exact items and price a merchant offered, so what the user approved is what gets charged
• Payment Mandate: authorizes a specific payment instrument against the approved cart
• Supports both 'human present' real-time purchases and 'human not present' delegated tasks (e.g. buy tickets the moment they go on sale) under the same mandate chain
• Payment-rail agnostic: designed to sit above card networks, real-time bank transfers, and stablecoin rails alike
• An A2A x402 extension, built with Coinbase, Ethereum Foundation and MetaMask, adds a production-ready path for agent-based crypto payments

Timeline

1. 2025-09-16Google announces the Agent Payments Protocol (AP2) with more than 60 payments and technology partners
2. 2026-04-28Google donates AP2 to the FIDO Alliance and releases spec version 0.2.0, and FIDO forms a Payments Technical Working Group chaired by Mastercard and Visa to develop it further as an open standard

Roadmap

Per Google's FIDO Alliance donation announcement, the stated direction is to make AP2 'remain platform-agnostic and community-led' under FIDO Alliance stewardship, working alongside Mastercard's Verifiable Intent within the same Payments Technical Working Group; FIDO states 'work has commenced within these workstreams' with no committed delivery dates yet.

Adoption signals

Announced by Google on 2025-09-16 with more than 60 payments and technology partners including Adyen, American Express, Coinbase, JCB, Mastercard, PayPal, Salesforce, ServiceNow and UnionPay International. On 2026-04-28 Google donated AP2 to the FIDO Alliance, which stood up a Payments Technical Working Group (chaired by Mastercard and Visa) to carry the spec forward as a multi-stakeholder standard alongside Mastercard's Verifiable Intent.

Competes against

Agentic Commerce Protocol

Governance is coded as corporate-led-open rather than open-foundation because, as of the FIDO donation announcement (2026-04-28), work inside FIDO's new working groups had just commenced and Google remains the visible primary steward of the public repo and docs. Distinct from the Agentic Commerce Protocol (OpenAI/Stripe) and Universal Commerce Protocol (Google): AP2 is the payment-authorization/mandate layer, not a checkout or catalog protocol.

Transport layer

REST/HTTP APIs defined via OpenAPI (Checkout, Delegated Payment and Product Feed specs); the 2026-04-17 revision added Model Context Protocol (MCP) compatibility alongside the existing REST bindings

Key features

• Agentic Checkout Spec: session-based checkout flow (create session, update, complete) that a merchant implements
• Delegated Payment Spec: lets a payment service provider issue a narrowly-scoped, single-use payment token to the agent instead of exposing raw card details
• Product Feed Spec: structured product data so an agent can index price and availability accurately
• Capability negotiation, payment handlers including seller-backed payment handlers, and an extensions mechanism for features like discounts
• Apache-2.0 licensed with dated spec releases (2025-09-29, 2025-12-12, 2026-01-16, 2026-01-30, 2026-04-17) tracked in a public repository
• Designed to keep the merchant as merchant-of-record throughout the transaction rather than routing purchases through the AI platform itself

Timeline

1. 2025-09-29OpenAI and Stripe launch the Agentic Commerce Protocol alongside ChatGPT's Instant Checkout, with Etsy as first live merchant
2. 2026-04-17Spec release adds cart, feed, orders, authentication and Model Context Protocol (MCP) compatibility

Roadmap

The repository states an intended 'path toward broader community governance' and 'neutral foundation stewardship as ecosystem matures,' though as of this research OpenAI and Stripe remain the named founding maintainers with no completed handoff.

Adoption signals

Launched 2025-09-29 alongside ChatGPT's Instant Checkout feature, with Etsy live on day one and additional Shopify merchants added afterward. OpenAI retired the Instant Checkout feature in March 2026, stating its initial version 'did not offer the level of flexibility that we aspire to provide,' shifting ChatGPT's own shopping experience toward dedicated retailer integrations. The ACP spec itself continued to be developed independently of that product decision, with a dated release on 2026-04-17 adding cart, feed, orders, authentication and MCP compatibility.

Competes against

Agent Payments ProtocolUniversal Commerce Protocol

Important nuance: 'Instant Checkout,' the ChatGPT feature, was retired in March 2026 (exact day not asserted; secondary sources disagree), but this is a product decision by OpenAI, not a deprecation of the ACP spec, which kept shipping dated releases afterward (most recently 2026-04-17, confirmed as the 'Latest Stable' release on the protocol's own GitHub repo) and is used by other integrators beyond ChatGPT. Status is therefore recorded as active at the spec level. Distinct from IBM's Agent Communication Protocol and Zed's Agent Client Protocol, which also use 'ACP' but operate in unrelated domains; this entry uses the full kebab-case name to avoid the collision.

Transport layer

UCP also supports multiple transports including A2A, MCP and APIs to provide flexibility for businesses and agents to communicate; capability discovery happens through a standardized JSON manifest published at /.well-known/ucp

Key features

• Capability-based architecture: businesses declare supported capabilities (Checkout, Identity Linking, Order, Payment Token Exchange) in a discoverable manifest rather than a single monolithic API
• Multiple transport bindings per capability: REST API, Agent2Agent (A2A), or Model Context Protocol (MCP), chosen per platform
• Compatible with AP2 for the payment-authorization leg of a transaction rather than defining its own payment-mandate scheme
• Extensions mechanism for augmenting capabilities with specialized functionality such as discounts
• Apache-2.0 licensed reference specification, SDKs and sample implementations maintained in a public GitHub organization
• Roadmap covers loyalty and member-benefit capabilities plus new verticals (food/restaurant discovery, hotel/lodging) beyond retail

Timeline

1. 2026-01-11Google announces the Universal Commerce Protocol at NRF 2026, co-developed with Shopify, Etsy, Wayfair, Target and Walmart
2. 2026-05-20Google expands UCP at Google Marketing Live with Universal Cart, buy-now-pay-later integrations, and new retail brand partners

Roadmap

The protocol's published roadmap page lists three priorities: deeper consumer-journey support (loyalty and member benefits, cross-sell/upsell), phased geographic expansion beyond the U.S. (India, Indonesia, Latin America and others), and new verticals (restaurant/food discovery, hotel/lodging capabilities), while explicitly stating this 'does not constitute a commitment to deliver specific features.'

Adoption signals

Announced by Google at NRF 2026 on 2026-01-11, co-developed with Shopify, Etsy, Wayfair, Target and Walmart and endorsed by more than 20 other companies including Adyen, American Express, Best Buy, Mastercard, Stripe and Visa. Expanded at Google Marketing Live on 2026-05-20 with a Universal Cart feature (save-and-checkout across retailers via Google Pay or the retailer's own checkout), buy-now-pay-later integration with Affirm and Klarna, and support for brands including Nike, Sephora and Target.

Competes against

Agentic Commerce Protocol

Scoped here as 'payments' because that is this lens's classification, but UCP is broader than payments alone; it also standardizes catalog discovery, cart and post-purchase support, with AP2 handling the payment-authorization mandate within a UCP-orchestrated flow. currentVersion is left null: no explicit version-numbering scheme was confirmed. Governance recorded as corporate-led-open since Google is the visible steward of the GitHub org, notwithstanding the wide co-developer list.

Transport layer

Native HTTP; payment terms and proofs are carried in the 402 response and companion request/response headers, with settlement executed on whichever underlying network (blockchain or, via newer schemes, card/bank rails) the chosen scheme specifies

Key features

• Repurposes HTTP 402 so payment requirements and proofs travel inside standard HTTP messages, no new handshake needed
• Scheme-based design separates transport, payment logic (exact, and proposed schemes like upto), and settlement network
• V2 (2025-2026) added multi-chain and multi-network routing, a plugin-driven SDK, and wallet-controlled sessions so repeat access can skip the full payment flow
• Automatic service discovery lets facilitators index x402-enabled endpoints from structured metadata
• Reference SDKs in TypeScript, Python and Go under Apache-2.0
• Designed to be network- and currency-agnostic rather than tied to one blockchain or issuer

Timeline

1. 2025-05-06Coinbase launches x402, an HTTP-402-based payments protocol, with partners including AWS, Anthropic, Circle and Cloudflare
2. 2025-09-23Coinbase and Cloudflare announce plans to launch an independent x402 Foundation for open governance of the protocol
3. 2026-07-14Linux Foundation announces the operational launch of the x402 Foundation as a 40-member, vendor-neutral governance body for the protocol

Roadmap

The x402 Foundation's charter, as stated in the Linux Foundation announcement, is to keep the protocol 'highly secure and adaptable, supporting multiple payment types, from traditional cards to stablecoins, without vendor lock-in,' with development now run as a collaborative, vendor-neutral process rather than a single-company roadmap.

Adoption signals

Originally incubated by Coinbase with launch partners including AWS, Anthropic, Circle and Cloudflare in May 2025. By the second half of 2025 it had processed a large and growing volume of transactions for paid APIs and AI-agent use cases. Governance was handed to an independent x402 Foundation, operationally launched under the Linux Foundation on 2026-07-14 with 40 founding members spanning payments and crypto infrastructure, including Adyen, AWS, American Express, Circle, Cloudflare, Coinbase, Fiserv, Google, Mastercard, Ripple, Shopify, Solana Foundation, Stellar Development Foundation, Stripe and Visa.

Not to be confused with Visa's or Mastercard's card-network agentic-commerce programs, which operate at a different layer (tokenized card credentials) rather than as an open HTTP-level protocol. Governance is coded here as open-foundation because the Linux Foundation operational launch (2026-07-14) is a completed handoff with a 40-member structure and a migrated canonical repo (x402-foundation/x402), independently confirmed. AP2 (Google) extends x402 as an optional crypto-payment rail rather than competing with it, so no competesAgainst entry is listed.

Transport layer

Protocol-agnostic; delegation and constraint data are carried as SD-JWT (selective-disclosure JWT) credentials meant to be embedded inside the payment/authorization flow of other protocols (AP2, ACP, UCP) rather than as a standalone transport

Key features

• Three-layer delegation model: identity issuer signs a credential, the user sets constraints, the acting agent proves it is operating within scope
• Eight constraint types (e.g. amount bounds, merchant allowlists, budget caps, recurrence terms) that are cryptographically bound and machine-verifiable
• Selective disclosure so an agent can prove only the constraints relevant to a given transaction, not the full credential
• Explicitly designed to interoperate with AP2, the Agentic Commerce Protocol, and the Universal Commerce Protocol rather than replace them
• Apache-2.0 licensed specification and reference implementation published on GitHub

Timeline

1. 2026-03-05Mastercard announces Verifiable Intent, an open-source authorization framework co-developed with Google
2. 2026-04-28Verifiable Intent is contributed to the FIDO Alliance's Payments Technical Working Group alongside Google's AP2

Roadmap

Per Mastercard's own reporting, the near-term plan is integration 'directly into Mastercard Agent Pay's intent APIs to support broader deployment with partners,' while standardization work continues inside the FIDO Alliance's Payments Technical Working Group.

Adoption signals

Announced 2026-03-05 by Mastercard as an open-source, standards-based framework co-developed with Google, with initial commitments from Fiserv, IBM, Checkout.com, Basis Theory and Getnet. On 2026-04-28 it was contributed, alongside Google's AP2, to the FIDO Alliance's new Payments Technical Working Group (chaired by Mastercard and Visa) to be developed as a shared industry standard.

This is a trust/authorization framework rather than a payment rail: it does not move money itself, similar in role to AP2's mandates but positioned as a cross-protocol, network-agnostic credential layer that AP2, ACP and UCP can each plug into. Not a competitor to AP2 (Google describes it as 'compatible with Agent Payments Protocol'). Distinct from Mastercard Agent Pay, a card-tokenization product, not an open spec, and therefore not tracked separately here. No source uses a short initialism for this protocol, so abbreviation is left null rather than repeating the name.

Transport layer

HTTP request signing per RFC 9421: Signature, Signature-Input, and Signature-Agent headers, with public keys published at a well-known directory URL referenced by the Signature Agent Card

Key features

• Cryptographic request signing built on the existing IETF standard RFC 9421 (HTTP Message Signatures)
• Signature-Agent header pointing to a machine-readable public-key directory
• Signature Agent Card: JSON metadata describing an agent's identity, purpose, and keys, reusing the OAuth Dynamic Client Registration namespace
• Bot-control vendors verify signatures against published keys and apply site owners' own access policies
• Transport- and vendor-neutral: implemented independently by Cloudflare, and consumed by non-Cloudflare infrastructure

Timeline

1. 2025-04-15draft-meunier-web-bot-auth-architecture -00 first submitted to the IETF
2. 2025-07-01Cloudflare integrates HTTP Message Signature verification into its Verified Bots Program
3. 2026-06-26draft-meunier-webbotauth-httpsig-protocol -00 submitted, splitting the signing protocol out of the original architecture draft
4. 2026-08-18draft-meunier-webbotauth-httpsig-protocol -02 submitted, the latest revision at verification time

Adoption signals

Cloudflare rolled cryptographic Message Signature verification into its Verified Bots Program starting July 1, 2025. Amazon Bedrock AgentCore Browser ships a preview 'Web Bot Auth' feature that signs outbound agent requests and lists Cloudflare, HUMAN Security, Akamai, DataDome, and F5 as supported verifying vendors, while flagging that the feature tracks a draft IETF spec still subject to change. As of the current draft text, the underlying documents remain individual Internet-Draft submissions, not yet adopted by an IETF working group.

Not an IETF-endorsed standard: the datatracker page explicitly states the draft 'is not endorsed by the IETF and has no formal standing in the standards process,' though production deployments already exist. Distinct from portable-identity-credential initiatives: Web Bot Auth authenticates the calling client/request via cryptographic signatures rather than issuing a portable identity credential for an agent.

Transport layer

Markdown file at /llms.txt: an optional byte-order mark, a mandatory H1 title, a blockquote summary, optional free-form markdown detail sections, and zero or more H2-delimited sections listing links (with optional notes) to further markdown resources

Key features

• Single well-known file path (/llms.txt) analogous to robots.txt
• Mandatory H1 heading plus a blockquote summary as the only required elements
• H2 sections list links to more detailed markdown files for deeper context
• Deliberately markdown rather than XML or JSON, to stay human- and LLM-readable while remaining fixed-format enough to parse programmatically
• v2 (2026) adds a standard way for agents to discover markdown versions of ordinary HTML pages

Timeline

1. 2024-09-03Jeremy Howard and Answer.AI publish the original /llms.txt proposal
2. 2026-08-10llms.txt v2 published, adding a standard mechanism for agents to discover markdown versions of pages

Adoption signals

Thousands of sites publish an llms.txt file; documentation platforms such as Mintlify generate one automatically for hosted sites. OpenAI, Anthropic, and Google's Gemini team publish their own llms.txt files for developer documentation. Version 2 of the spec was published August 10, 2026, the first revision since the original 2024 proposal.

Scope tension, recorded rather than resolved silently: this is a static, crawlable file at a well-known path with no live request/response exchange, unlike every other entry in this dataset (which define a message format traded between defined roles over a live session). It is included because it is a strict, agent-specific, machine-parseable format (not a plain web standard, and not a loose convention), but it does not itself define an interaction protocol the way MCP, A2A, or ACP do. Distinct from NLWeb (static curated file vs. live natural-language query endpoint).

Transport layer

HTTP REST endpoints (POST /ask, with an /mcp endpoint for MCP clients) exchanging JSON requests and Schema.org-vocabulary JSON responses; supports streaming and list/summarize/generate response modes

Key features

• /ask endpoint: natural-language query in, structured JSON answer out
• Responses expressed in Schema.org vocabulary already used by many sites' structured data
• Every instance doubles as an MCP server for MCP-based agent clients
• Supports list, summarize, and generate response modes plus conversational context via a 'prev' parameter
• Technology-agnostic reference implementation: site owners choose their own backend data source and model
• Open-sourced under the MIT license

Timeline

1. 2025-05-19Microsoft announces NLWeb at Build 2025 with a named cohort of publishing and ecosystem collaborators

Adoption signals

Announced by Microsoft at Build 2025 (May 19, 2025). Initial publishing and ecosystem collaborators named at launch included Chicago Public Media, Common Sense Media, DDM (Allrecipes/Serious Eats), Eventbrite, Hearst (Delish), Inception Labs, Milvus, O'Reilly Media, Qdrant, Shopify, Snowflake, and Tripadvisor. The reference implementation has since moved to a community-run GitHub organization (nlweb-ai) with ports appearing for .NET and WordPress.

Complementary to, not competing with, model-context-protocol: an NLWeb instance is itself an MCP server, so the two are layered rather than rival standards. The repo transfer from microsoft/NLWeb to nlweb-ai/NLWeb is now confirmed directly (the GitHub API resolves microsoft/NLWeb to the nlweb-ai/NLWeb canonical record), so this is a confirmed migration rather than an ambiguous one.