Model Context Protocol (MCP)
The open protocol that connects AI applications to tools and data: how it is built, how to implement it and how to test it.
14 articles in reading order
- What is MCP (Model Context Protocol)MCP is an open protocol that connects AI applications to tools and data. Architecture, primitives, transports, the 2026-07-28 specification and security.
- Tools in MCPAn MCP tool is a function a server exposes for a model to call. Definition fields, tools/list and tools/call, results, errors, names and x-mcp-header.
- Resources and resource templates in MCPAn MCP resource is data a server exposes by URI as context. URI schemes, resources/list and resources/read, templates, contents, annotations and subscriptions.
- Prompts in MCPAn MCP prompt is a message template a server offers and the user invokes. prompts/list and prompts/get, arguments, completion, content types, comparison.
- MCP transports: stdio and Streamable HTTPAn MCP transport defines how JSON-RPC messages are framed and delivered. stdio, Streamable HTTP, headers, status codes, history and backward compatibility.
- Stateless MCP: changes in 2026-07-28Stateless MCP is the request model of the 2026-07-28 specification: per-request _meta, server/discover, MRTR, state handles, caching and migration.
- Authorization in MCPAuthorization in MCP is an OAuth 2.1 based flow for HTTP servers: bearer tokens, resource metadata, client registration, scopes and issuer checks.
- MCP securityMCP security covers the risks of connecting AI applications to MCP servers: tool poisoning, confused deputy, token passthrough, SSRF, local servers, CVEs.
- MCP extensionsMCP extensions are optional features outside the core protocol: identifiers, negotiation, lifecycle, Tasks, MCP Apps, Enterprise-Managed Authorization, Skills.
- MCP server over an existing APIAn MCP server over an existing API adapts it for a model: choice of tools, descriptions, parameters, response size, errors, authorization and state.
- Testing MCP servers: the basicsHow to test an MCP server: test levels, protocol conformance, MCP Inspector and the conformance suite, security checks and evaluation with models.
- Evaluating an MCP server with modelsEvaluation of an MCP server: tasks with verifiable outcomes, repeated trials, pass@k and pass^k, metrics, the model and host matrix, transcript analysis.
- Regression testing of an MCP serverRegression testing of an MCP server: snapshots of tools/list, breaking changes, protocol version support, model behavior after description edits, CI.
- Publishing an MCP server to the MCP RegistryThe MCP Registry is the official catalog of MCP server metadata. The server.json format, namespaces, ownership checks, mcp-publisher, versions, subregistries.