MCP: The USB-C for AI Tools
The Model Context Protocol is an open standard for connecting AI applications to external data and tools — and it has become the de-facto industry protocol since its November 2024 launch.
What changed
Anthropic launched MCP in November 2024 as an open-source protocol based on JSON-RPC 2.0, taking design inspiration from the Language Server Protocol. MCP defines three primitives: Tools (executable functions the model can call), Resources (structured data included in context), and Prompts (reusable templates). The community has built thousands of MCP servers since launch, SDKs exist for all major languages, and the industry has broadly adopted MCP as the standard for connecting agents to tools. Anthropic also launched an MCP Connector on the API that lets you connect to remote MCP servers directly from the Messages API without a separate client. MCP Apps (SEP-1865) was formalized in early 2026, extending the protocol to support interactive React-based UIs from MCP servers.
Why it matters
MCP eliminates the bespoke integration tax. Before MCP, every AI tool needed its own custom connector to every data source. With MCP, build one server per data source and any MCP-compatible host can connect. This compounds: as more hosts adopt MCP (Claude, IDEs, custom agents), your MCP server works everywhere automatically. The API connector removes the need to run a local MCP client, making server-side agent architectures much simpler.
What can be built
- MCP servers exposing internal APIs, databases, or file systems to Claude agents
- Developer tools that read live codebase context via MCP resources
- Multi-agent workflows where each agent has a dedicated MCP server for its domain
- Interactive dashboards delivered from MCP servers using MCP Apps (SEP-1865)
- MCP servers wrapped around legacy REST APIs to make them agent-accessible
- Local MCP development servers for testing agent capabilities before deploying
Limitations & risks
- MCP does not yet have a server discovery protocol — server URLs must be known in advance
- The OAuth authorization spec for MCP (June 2025 update) is still evolving — review before implementing auth
- MCP Apps (SEP-1865) is an early 2026 formalization — expect iteration on the spec
- Running untrusted MCP servers can expose your agent to prompt injection via malicious tool responses
- MCP server performance directly impacts agent latency — slow resource reads block the agent
- No built-in rate limiting in the MCP spec — implement at the server layer