Exposes your tools.
A server wraps your data source or API behind a typed tool contract: descriptions, input schemas, and outputs any MCP client can call.
Model Context Protocol
MCP Is the USB-C Port for AI Applications. Here Is How to Build With It.
The Model Context Protocol is an open standard that connects AI applications to data sources, tools, and workflows - so agents can read your calendar, query your databases, and drive your integrations. Build once, integrate everywhere: Claude, ChatGPT, Cursor, and VS Code all speak MCP.
Last updated September 17, 2026. Covers the MCP specification dated 2026-07-28. Verify current spec details at modelcontextprotocol.io.
Definition
One protocol between AI apps and every external system.
MCP is an open-source standard for connecting AI applications to external systems. AI applications like Claude or ChatGPT use it to reach local files, databases, search engines, calculators, and specialized prompts - the same way USB-C standardizes how devices connect to peripherals. The current specification (2026-07-28) adds a public registry and refined OAuth-based authorization, and MCP Apps let developers ship interactive UIs that run inside AI clients.
What MCP enables
- Agents that read your calendar, Notion, and files
- Claude Code generating a web app from a Figma design
- Enterprise chatbots querying multiple databases safely
- Models driving 3D tools, printers, and real hardware
- Interactive MCP Apps running inside AI clients
Why it matters: developers cut integration time because one MCP server serves every MCP-compatible client; agents gain a whole ecosystem of tools without custom glue code; and users get assistants that can act, not just chat. Claude, ChatGPT, VS Code, Cursor, and MCPJam all support the protocol today.
Architecture
Servers, clients, and the registry.
Lives in the AI app.
Claude, ChatGPT, Cursor, and VS Code ship MCP clients that discover a server's tools and route model requests to them.
Stdio or HTTP.
Local servers run over stdio for privacy; remote servers use streamable HTTP with OAuth so teams can share tools securely.
Discovery for the ecosystem.
The MCP registry indexes public servers so builders can find and verify tools instead of reinventing integrations.
Build
Ship your first MCP server in an afternoon.
Start with one narrow, read-only tool. A weather lookup, a database reader, or a search wrapper teaches the full contract - tool descriptions, schemas, errors - without any destructive risk. Official SDKs exist for TypeScript and Python, and the Inspector app helps you debug tool calls end to end.
MCP server checklist
- Pick one narrow tool with a clear input schema.
- Write tool descriptions the model can act on without examples.
- Validate every input; never trust model-generated parameters.
- Return structured errors the client can display.
- Test in the MCP Inspector before connecting a real client.
- Add rate limits and logging before sharing the server.
Security rule: an MCP server is remote code execution with a friendly interface. Treat every model-initiated tool call like an untrusted user request.
Guardrails
Connect tools without creating an attack path.
Scope every credential.
Read-only database users, narrow API scopes, and per-server secrets. Never give one server a god token.
Distrust tool output.
Data retrieved by MCP tools can contain instructions. Keep tool results in structured channels, never concatenated into system prompts.
Gate destructive calls.
Writes, deletes, payments, and emails need a human approval step or a sandboxed dry-run mode.
People also ask
MCP questions builders ask.
What is the Model Context Protocol in simple terms?
MCP is an open standard that lets AI applications connect to external tools and data - like a USB-C port for AI apps. Build one MCP server and any MCP-compatible client, from Claude to VS Code, can use your tools.
Is MCP better than custom function calling?
Custom function calling ties integrations to one model's API. MCP is model-agnostic: the same server works across Claude, ChatGPT, Cursor, and VS Code, and the registry plus OAuth support make sharing tools easier. Use native function calling for single-model prototypes and MCP when integrations should outlive one provider.
Do MCP calls cost extra tokens?
Tool descriptions and results add tokens to the context, so costs rise with tool count. Keep tool descriptions tight, return compact structured results, and follow agent cost discipline - parallel sub-agent sprawl is a known token trap.
Which AI clients support MCP?
Claude, ChatGPT, VS Code, Cursor, and MCPJam all support MCP today, and the ecosystem list keeps growing. Check the official clients page for the current roster.