Model Context Protocol (MCP)
Model Context Protocol (MCP) is an open standard for connecting AI assistants like Claude and ChatGPT to external data sources, tools, and systems. It provides a unified way for AI applications to read files, query databases, call APIs, and perform actions through a standardized protocol.
itArtificial intelligence and machine learning | OpenSkills.info
Course pathWalk it in order
Look it upDip in anytime
Go furtherLeaves this page
Don't Panic
Don't Panic — Model Context Protocol (MCP)
Model Context Protocol, or MCP, is the agreed way for an AI application to talk to outside systems. It does not make the model wiser, and it does not replace an API. It gives the model a standard route to live context and actions, which is considerably less mystical than it sounds and considerably more useful than copying information between tabs.
Before MCP, every AI application needed its own special adapter for every database, file store, calendar, or development tool. That is the M times N arrangement: every new application meets every new system and they all exchange introductions. MCP changes the arithmetic to M plus N: build a server for the system, and compatible applications can use that server. The protocol is the shared plug shape, not the device at either end.
The cast has three members. A host is the AI application. Inside it, a client keeps one connection to one server. The server offers the useful things: tools for actions, resources for read-only context, and prompts for reusable interaction templates. One host can have several clients, which is why an assistant can consult files, a database, and an API without turning into a small box of incompatible cables.
The useful distinction is between a tool and a resource. A tool can query a database, send a message, or alter the outside world. A resource supplies data without that side effect. That sounds like vocabulary homework until a tool has permission to do something expensive. Then it becomes the difference between looking through a window and being handed the keys.
For a local server, STDIO transport carries protocol traffic through standard input and output. For a remote service, Streamable HTTP transport uses HTTP and can use authentication. Under either transport, JSON-RPC messages carry discovery, requests, responses, and notifications. A client first learns what a server offers, then lists and calls the tools it actually needs.
Read the Introduction for the architecture and the places MCP fits. Use Slides for the map of hosts, clients, servers, transports, and primitives. Keep the Cheatsheet nearby when names such as tools/list and tools/call start appearing in configuration or code. The Practice Reference is for making a server connection work. Field Notes is for the part that tends to arrive after the first successful connection: deciding what the server should be allowed to do.
Where this skill leads
Relevant careers
See how this topic contributes to broader role-level skill maps.
Sources
- https://modelcontextprotocol.io/introduction
Supports
- MCP is an open protocol that standardizes connections between AI applications and external context providers.
- https://modelcontextprotocol.io/specification/2025-11-25
Supports
- The MCP specification defines hosts, clients, servers, capability negotiation, and protocol messages.
- https://modelcontextprotocol.io/docs/2026-07-28/learn/architecture
Supports
- MCP uses a host-client-server architecture to isolate tool and data integrations from the AI application.
- https://developers.openai.com/codex/mcp
Supports
- Codex can be configured to use MCP servers as external tool providers.
- https://docs.anthropic.com/en/docs/agents-and-tools/mcp
Supports
- Anthropic documents MCP as a way to connect Claude applications to external tools and data.
- https://github.com/modelcontextprotocol/servers
Supports
- The repository lists reference and community MCP server implementations.
- https://github.com/modelcontextprotocol/inspector
Supports
- MCP Inspector is a developer tool for testing and debugging MCP servers.
- https://code.visualstudio.com/docs/copilot/chat/mcp-servers
Supports
- Visual Studio Code documents MCP server configuration for Copilot Chat.
- https://ts.sdk.modelcontextprotocol.io/v2/get-started/first-client.html
Supports
- A local STDIO client can spawn a server, list its tools, and invoke a tool with schema-valid arguments.
- https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization
Supports
- Remote MCP authorization uses OAuth resource metadata, audience-bound access tokens, and secure token handling requirements.
- https://owasp.org/www-chapter-stuttgart/assets/slides/2025-09-25_All_About_MCP_Security.pdf
Supports
- MCP server tool definitions can change after approval and require source review, change monitoring, and restrained tool approval.
- https://lenshq.io/blog/mcp-security-tool-poisoning-threat-model/
Supports
- Tool descriptions, parameters, errors, and outputs can be used to influence model behavior in tool-poisoning attacks.
- https://code.claude.com/docs/en/mcp
Supports
- Claude Code supports connecting tools through MCP.
- https://help.openai.com/en/articles/11487775-connectors-in-chatgpt
Supports
- ChatGPT supports connected apps and external service integrations.
- https://cursor.com/docs
Supports
- Cursor documents MCP configuration for external tools.
- https://docs.github.com/en/copilot/how-tos/provide-context/use-mcp-in-your-ide/extend-copilot-chat-with-mcp
Supports
- GitHub Copilot can configure MCP servers and limit access to approved scopes.
