Model Context Protocol

How AI connects
to tools and data.

Model Context Protocol gives AI applications a common interface for calling tools, reading data, and retrieving reusable prompts from servers.

Revision 2026-07-28

Host, client and server
AI workspaceMCP host
MCP clientProtocol messages
Workspace serverMCP server
ToolsResourcesPrompts

Overview

Three ways a server can help.

An MCP host is the AI application that manages the user experience and the model. Within it, MCP clients communicate with servers that expose specific capabilities. The host controls which servers are available and what context or operations the model can use. [1]

tools/list · tools/call

Tools

Callable operations such as calculating a metric or looking up a document. A tool has a name and an input schema, and can return content plus structured data. The host decides how to expose the tool and when a call needs approval. [2]

resources/list · resources/read

Resources

Context that a server exposes at a URI, such as a dataset, document, or schema. An application can select what to read and decide how to present or include it. Resource templates can describe URIs that take parameters. [3]

prompts/list · prompts/get

Prompts

Reusable templates that produce messages for a model. A client can retrieve a named prompt and supply arguments such as a topic or document title. The application decides how users find and apply the template. [4]

Interactive example

Inspect a server’s interface.

Choose a request a research workspace could issue to a fictional MCP server. Inspect the JSON-RPC body and the metadata that identifies the 2026-07-28 revision. The examples are local to this page.

Protocol explorerIllustrative · Local only

Discover available tools

A tools/list request asks for the available tool descriptions and schemas. A server can paginate its catalog. A client should check the schema for the particular tool before building its call.

tools-list.json
{  "jsonrpc": "2.0",  "id": 1,  "method": "tools/list",  "params": {    "_meta": {      "io.modelcontextprotocol/protocolVersion": "2026-07-28",      "io.modelcontextprotocol/clientInfo": {        "name": "Research Workspace",        "version": "0.1.0"      },      "io.modelcontextprotocol/clientCapabilities": {}    }  }}

Implementation

How MCP works on the wire.

Separate the application roles, the request format, the transport, and authorization. Each answers a different implementation question.

Match the protocol revision

MCP uses JSON-RPC 2.0. The 2026-07-28 core is stateless at the protocol level, with the revision and relevant client capabilities carried in request metadata. A client may call server/discover to learn supported revisions and capabilities. An application can still keep its own state and refer to it through explicit identifiers.

In 2025-11-25 and earlier session-based revisions, initialize established the version and capabilities. Integrations must account for the lifecycle the client and server actually support. [1][5]

Choose a transport

With stdio, a local client typically starts the server process and exchanges messages over standard input and output. Streamable HTTP carries MCP between a client and an HTTP endpoint, including streams where the interaction requires them. The JSON in the explorer shows the body. HTTP implementations also follow the binding’s header and framing rules. [1][5]

Handle results and additional input

Tool results can include structured content alongside user-facing content. An execution error may be returned inside the tool result with isError set. Errors such as an invalid method are protocol-level JSON-RPC errors. Handle these outcomes separately.

In the 2026-07-28 revision a supported interaction can ask for more input and continue across multiple requests. The client supplies inputResponses and echoes requestState when the server provided it. [1][2]

Enforce authorization at the boundary

The MCP authorization framework defines how an HTTP server can protect access with OAuth. Apply authorization to the intended resource and operation, and keep local process permissions in mind for stdio. A tool catalog describes capabilities. The host and server still make access decisions and can require approval before executing an operation with side effects. [2][6]

The Agentlithic architecture

Where MCP fits.

In the Agentlithic architecture, MCP is the interface an AI application or agent could use to reach authorized tools, documents, and prompt templates. A2A handles the conversation with another agent when the work is delegated. The two can appear in the same workflow at different boundaries.

Companion guideA2AAgent2Agent Protocol

Sources

Revision 2026-07-28
  1. Model Context Protocol Specification, 2026-07-28

    Model Context Protocol project · Architecture, Base Protocol Overview, Versioning and Compatibility, and Message Patterns

  2. Model Context Protocol Specification, 2026-07-28

    Model Context Protocol project · Server Features → Tools, including Tool Results and Error Handling

  3. Model Context Protocol Specification, 2026-07-28

    Model Context Protocol project · Server Features → Resources, including Resource Templates

  4. Model Context Protocol Specification, 2026-07-28

    Model Context Protocol project · Server Features → Prompts, including Listing and Getting Prompts

  5. Model Context Protocol Specification, 2026-07-28

    Model Context Protocol project · Transports → Overview, stdio, and Streamable HTTP. See also the maintainers’ release article, “The 2026-07-28 Specification.”

  6. Model Context Protocol Specification, 2026-07-28

    Model Context Protocol project · Authorization and Security Best Practices