In 2026, Claude developers have two distinct ways to give Claude access to external capabilities: built-in tool calling (define a schema, Claude calls it, you execute) and Model Context Protocol (MCP) servers (a standardized protocol for external capability servers). Both work; they solve overlapping problems differently.
Tool calling — single-app capabilities
Define a JSON-schema for each function your code can execute.
Pass the tool list with each request.
Claude returns structured tool-call requests; you execute and feed results back.
Lives inside your app — bespoke per integration.
Use when: the tool is specific to your app, you control both sides of the call, and you don't need other agents to use it.
MCP — reusable capability servers
Stand up an MCP server that exposes capabilities over a standardized protocol.
Any MCP-aware client (Claude Desktop, Claude Code, Cursor, etc.) can connect and use it.
Capabilities are advertised via a manifest; the client decides which to invoke.
Lives outside any single app — shared infrastructure.
Use when: the capability is reusable across multiple apps or agents, you want users of any MCP-aware client to benefit, or you're building a third-party tool ecosystem.
Decision matrix
Question
If yes
If no
Will multiple agents need this capability?
MCP
Tools
Do you need real-time per-request control?
Tools
Either
Is the tool definition private to your app?
Tools
Either
Are you building dev infrastructure?
MCP
Tools
The hybrid pattern
Many production agents use both: in-process tools for app-specific actions (your database, your business logic), plus MCP connections for general-purpose capabilities (filesystem, browser, search). The tool list passed in each request includes both.
Browse the Claude Skills Browser for reusable capabilities Anthropic ships, several of which are exposed as MCP servers.
Frequently asked questions
Are tool calls and MCP mutually exclusive?
No. A single Claude request can include both in-process tool definitions and MCP-connected capabilities. The client merges them into one tool list. Production agents commonly use this hybrid pattern.
Do tool calls cost more than regular API calls?
Tool definitions count as input tokens, so larger tool lists cost more per request. Cache the tool definitions if you reuse the same set — they become eligible for the 90%-off cached-read pricing.
Can MCP servers be private to my organization?
Yes. MCP is a protocol, not a directory. You can run an MCP server inside your private network and configure your MCP-aware clients to connect to it. No public listing required.