Aliases: Model Context Protocol
MCP (Model Context Protocol)
An open standard for connecting AI models to tools and data sources, replacing per-app custom integrations with one reusable interface.
Last reviewed: July 25, 2026
What is MCP?
MCP standardizes how AI agents connect to external systems. An MCP server wraps a data source or service (GitHub, Postgres, Slack, your internal API) and exposes three primitives — tools (actions the model can call), resources (data it can read), and prompts (reusable templates). Any MCP-compatible client (Claude, IDEs, agent frameworks) can then use it without custom integration code.
Why it took over
Before MCP, connecting M applications to N services meant M×N bespoke integrations. MCP makes it M+N — write one server for your service and every AI client can use it. Introduced by Anthropic in late 2024 and adopted across major vendors, by 2026 it’s the de facto tool-connectivity layer of the agent stack: “agent integrations” went from custom plumbing to a portable interface.
The USB-C analogy
MCP is to AI tools what USB-C is to peripherals: a standard port. It doesn’t make tools good — it makes them pluggable. Tool design quality (clear descriptions, tight schemas, helpful errors) still determines whether the agent uses them well.
Security reality — read this part
MCP moves the trust boundary. An MCP server runs with its credentials, and the model decides when to call it. Real risks seen in the wild: prompt injection (malicious content in a fetched document instructs the agent to exfiltrate data via another tool), overly broad OAuth scopes, and unvetted third-party servers. Treat MCP servers like production dependencies: least-privilege credentials, read-only where possible, human confirmation for destructive actions, and audit logs — the same zero-trust thinking you’d apply to any service account.
What people get wrong
- Exposing 50 tools at once. Models choose worse as tool count grows; curate a small set per use case.
- Thin wrappers around raw REST APIs. Good MCP tools are task-shaped (“create_invoice_with_line_items”), not endpoint-shaped (“POST /v2/resources”).
- Skipping auth review. A server with your admin token turns any prompt-injection into an admin action.
The Problem MCP Solves
Before MCP, connecting an AI model to an external tool or data source — a database, a file system, a SaaS API — typically required building a custom, one-off integration specific to that combination of model and tool, meaning the same integration effort had to be repeated for every new model or application that wanted to use the same tool. MCP defines a standard protocol that any compliant tool or data source can implement once, and any compliant AI application can then connect to, similar in spirit to how a USB standard lets any compliant peripheral connect to any compliant computer without a custom driver for every specific pairing. Introduced by Anthropic in late 2024 and since adopted more broadly across the AI industry, MCP has become a common way for AI applications to gain access to external tools, file systems, databases, and APIs through a consistent, reusable interface rather than bespoke integration code for each new capability.
Historical figures and technical concepts for informational purposes only. Not technical, professional, legal, or financial advice. Sources: Official Documentation.