What is MCP?
MCP is a protocol that lets AI agents talk to external tools and resources in a consistent way. Servers expose tools and resources; clients (agents) can discover and call them.
Model Context Protocol & Agents
A developer-friendly primer on building interoperable tools and UI surfaces for AI agents. Learn the essentials of MCP servers, tools, and widgets—by example.
Tools
Tools are typed operations an agent can call—for example, “add_todo” or “complete_todo”. They return human-readable content and structured data the UI can render.
Widgets
Widgets are small UI views attached to tool output. They render structured results, stay in sync with updates, and can call tools directly when permitted.
Why it matters
MCP standardizes how agents integrate with systems. It reduces bespoke glue code, improves safety via schemas, and enables richer experiences with minimal friction.
How MCP Works
MCP establishes a consistent contract between agents and external systems. Servers expose tools and resources; agents discover, validate, and invoke them using schemas and well-defined protocol messages.
Discovery
Agents connect and list available capabilities: tools (operations) and resources (data or templates). Clear metadata makes orchestration predictable and safe.
Invocation
Inputs are validated against schemas; outputs include human-readable content and structured data. This duality ensures clarity for both users and UIs.
Templating
Widgets render structured results from tools. They stay synchronized with agent updates and can call permitted tools directly to continue workflows.
Transport
Servers typically speak HTTP(S) for portability. Small, stateless exchanges keep integrations simple while remaining robust under scale.
Design to be composable: prefer small tools with clear boundaries over monolithic actions. Emphasize validation and explicit responses for reliability.
Agents in Practice
- Planning: Agents create a plan, then select tools based on schemas, permissions, and context.
- Execution: Tool calls are batched or sequenced. Each returns content and structured data shaping subsequent steps.
- Feedback loops: Results update the agent’s state and, when widgets are present, the user’s view.
- Fallbacks: On validation failures, agents reformulate inputs or present clear error messages.
The goal is predictability: schemas, consistent outputs, and minimal side effects help agents reason and users stay informed.
Security & Safety
- Schema validation: Enforce input constraints to prevent unsafe or ambiguous calls.
- Least privilege: Only expose necessary tools; gate sensitive operations.
- Auditability: Keep logs of tool invocations and results for traceability.
- Determinism: Favor idempotent actions and explicit responses to reduce state drift.
Design Principles
- Schema-first: validate inputs and outputs for safety.
- Small, composable tools over monoliths.
- State minimalism: prefer idempotent operations and explicit responses.
- UI optionality: widgets enrich results without coupling agents to UIs.