MCP vs A2A vs Agent Skills: what each one does

MCP is an open protocol for connecting LLM applications to external data sources and tools. It uses JSON-RPC 2.0 messages between hosts (the AI application), clients (connectors inside the host) and servers (programs that provide context). The latest specification revision is dated 2026-07-28.

MCP servers expose three core primitives: tools (functions the model can execute), resources (context data) and prompts (reusable templated messages). Clients can offer elicitation, which lets a server ask the user for more input. MCP's scope is context exchange only: it does not dictate how AI applications use LLMs or manage the provided context.

As of this revision MCP is described as stateless. Every request carries the protocol version and relevant capabilities in its _meta field, and servers advertise versions and capabilities through the mandatory server/discover request. The sampling and logging client primitives are marked deprecated as of protocol version 2026-07-28.

The specification lists optional extensions, including Tasks (async long-running operations), MCP Apps (interactive UI rendered inline) and 'Skills over MCP' (structured instructions for agent workflows, discovered and consumed through MCP). On governance, code and spec contributions use the Apache License 2.0, and changes are proposed as Specification Enhancement Proposals (SEPs).

A2A is an open standard for communication between AI agents. It gives agents from different frameworks and vendors a shared language. It was originally developed by Google and donated to the Linux Foundation, and its license is Apache 2.0.

A2A treats agents as opaque peers. They collaborate without exposing internal logic, memory or proprietary tools, and the A2A page argues that wrapping agents as tools limits what they can do. Core operations include SendMessage (single response) and SendStreamingMessage (streaming task updates and artifacts), with support for long-running work through streaming and asynchronous execution.

A server publishes an Agent Card, typically at /.well-known/agent-card.json. Clients fetch it to learn capabilities and endpoint, and to see security schemes that indicate whether a token is needed. In the 1.0.0 spec an AgentCard requires name, description, supportedInterfaces, version, capabilities and skills. Discovery is via the well-known URI, registries/catalogs, or direct configuration.

The latest released A2A specification version is 1.0.0 (previous: 0.3.0, 0.2.6, 0.1.0). It defines three protocol bindings (JSON-RPC, gRPC, HTTP+JSON/REST), with a Protocol Buffers data model as the normative source. The v1.0 changes include renaming methods (message/send became SendMessage), merging part types into one Part type, removing implicit and password OAuth flows, and adding Device Code flow and a pkceRequired field.

Agent Skills are a lightweight open format. A skill is a folder containing a SKILL.md file with metadata (name and description at minimum) and instructions, optionally with scripts/, references/ and assets/ directories. The format was originally developed by Anthropic, released as an open standard and adopted by a growing number of agent products. Development is on GitHub (agentskills/agentskills).

Skills use progressive disclosure in three stages: discovery (only name and description loaded at startup), activation (full SKILL.md read when a task matches), and execution (bundled scripts or referenced files loaded as needed). The spec's size guidance: metadata is about 100 tokens per skill at startup, the body is recommended under 5000 tokens, and SKILL.md should stay under 500 lines with detail moved to separate files.

Per the open spec, name and description are the required frontmatter fields. The name is up to 64 characters, lowercase letters, numbers and hyphens, and must match the directory. The description is up to 1024 characters and should say what the skill does and when to use it. Optional fields are license, compatibility, metadata and allowed-tools (experimental).

Claude Code skills follow the open standard and add extras such as invocation control, subagent execution (context: fork) and dynamic context injection. Its doc says all frontmatter fields are optional, with description recommended. It locates skills at enterprise, personal, project, nested subdirectory and plugin levels; a project skill lives at .claude/skills/<name>/SKILL.md. Custom commands under .claude/commands have been merged into skills.

The A2A comparison page frames MCP as agent-to-tool, a 'vertical' direction that deepens one agent. It frames A2A as agent-to-agent, a 'horizontal' direction across teams or organizations. It calls the two complementary.

The A2A site's own example has a shop-manager agent talk to mechanic agents over A2A. A mechanic uses MCP to call tools like a diagnostic scanner, and A2A again to query a parts-supplier agent.

The MCP specification also lists 'Skills over MCP' as an optional extension. It describes it as structured instructions for agent workflows, discovered and consumed through MCP.

This page does not rank these options and does not tell you which to choose. It only repeats how each project describes itself.

The MCP specification describes an open protocol for connecting LLM applications to external data sources and tools. The A2A home page describes an open standard for communication between AI agents. The Agent Skills overview describes a lightweight open format: a folder containing a SKILL.md file with metadata and instructions.

The A2A comparison page calls MCP and A2A complementary. Its own example uses both: a mechanic agent uses MCP to call tools like a diagnostic scanner, and A2A to query a parts-supplier agent.