Model Context Protocol (MCP)
Model Context Protocol standardizes how AI applications discover and exchange tools, resources, and reusable prompts with external servers.
Download Markdown (.md)Soroush’s Product Wiki · https://wiki.soroush.uk/topic/model-context-protocol/
01What it means
Model Context Protocol, or MCP, defines communication between a host application's clients and servers exposing capabilities. A host may connect to multiple servers; each client manages a connection. Core capabilities include callable tools, readable resources, and prompt templates. MCP standardizes the integration boundary, while the host still decides how capabilities are presented, approved, and used by its model.[1][2]
- Host applicationUser experience, model, policy
- MCP clientConnection and capability negotiation
- MCP serverTools, resources, prompts
Structured capability results and content
Servers expose tools, resources, and prompts; the host controls presentation, approval, authorization, and model use.
02Why it matters
A shared interface can reduce custom integration work and make capabilities reusable across compatible applications. It also introduces trust and permission decisions: a discoverable tool is not automatically appropriate for a user or safe for a model to invoke.
03When to use it
Consider MCP when several AI applications need consistent access to external capabilities or an integration must work across compatible hosts. A direct API can be simpler for a single fixed workflow. Check actual client support, transport, authorization, and protocol-version compatibility before promising interoperability.
04How it works
- Define the user tasks, data boundaries, and external capabilities the integration should expose.
- Implement or select a server with narrow tools, resource access, and documented capability schemas.
- Connect through the host's client and verify initialization, supported capabilities, and compatible protocol versions.
- Enforce authorization and input validation at the service boundary, with user review for consequential actions.
- Test untrusted server content, failures, revoked access, and audit records before rolling out the integration.
05A practical example
A product team exposes issue lookup and draft creation through an MCP server used by two internal assistants. Each user retains their existing issue permissions. Creating a draft is separate from publishing it, and the host shows the intended change for approval. Standardized discovery reduces duplicate integration code without replacing the service's access checks.
06Where it breaks
Host implementations differ and protocol versions evolve. Interoperability does not guarantee equivalent user experience, trustworthy server descriptions, or correct model behavior. External results can contain prompt injection, and authentication alone does not establish permission for every action.
Common mistakes
- Calling MCP a model or an agent rather than an integration protocol.
- Assuming every compatible host supports every server capability.
- Treating tool discovery as authorization to execute it.
- Passing credentials through without validating their intended audience and scope.
07Sources
- Architecture overviewModel Context Protocol · Official documentationArchitecture overview
- Security Best PracticesModel Context Protocol · Official documentationSecurity Best Practices