Tools & retrieval
What is MCP?
Also known as: Model Context Protocol
Model Context Protocol (MCP) is an open standard for connecting agents to tools and context servers.
MCP standardizes discovery and invocation. Your app still owns auth, guardrails, and logging.
An MCP server exposes `list_issues` and `create_issue`. The host agent discovers tools at runtime over the MCP protocol instead of hard-coding GitHub HTTP calls. Adding a new tool server does not require changing the agent code.
Visual
MCP architecture
Agent host connects to multiple MCP servers, each providing different tools.
Why it matters
Model Context Protocol (MCP) standardizes how agents connect to tools. Before MCP, every integration required custom code — parsing specific API responses, handling auth differently, managing errors uniquely. MCP provides a shared contract for tool discovery, invocation, and result format.
MCP matters because it makes agent tool sets composable. You can swap tool servers, add new capabilities, and share integrations without rewriting agent logic.
How MCP works
An MCP client (the agent host) connects to MCP servers (tool providers). The server advertises its available tools with schemas. The client presents these tools to the model. When the model requests a tool call, the client routes it to the correct server, which executes and returns the result.
The protocol handles transport, serialization, and error reporting. Your application still owns authentication, rate limiting, and guardrails.
MCP vs. direct API integration
Direct API integration gives you full control but requires custom code per service. MCP gives you standardized discovery and invocation but requires the tool provider to support the protocol.
In practice, many teams use MCP for external services and direct integration for internal tools. The two approaches complement each other.
Key takeaways
- 1MCP standardizes how agents discover and call tools.
- 2The protocol handles transport; your app handles auth and guardrails.
- 3MCP makes tool sets composable — add servers without changing agent code.
- 4MCP is a protocol, not a framework or agent system.
Common mistakes
- ✕Thinking MCP replaces the need for guardrails — more tools means more side effects to gate.
- ✕Exposing MCP servers without authentication or rate limiting.
- ✕Confusing MCP with an agent framework — it is a tool integration protocol.
In practice
Most MCP deployments start with 2–3 servers: a code/repository server, a search server, and a domain-specific server. The agent discovers available tools at startup, and the model chooses which to call based on the task. Adding a new server immediately extends the agent's capabilities without code changes.
Related terms
Concept neighborhood
Terms linked from MCP in the glossary graph.
FAQ
- Is MCP an agent framework?
- No. It is a protocol for connecting agents to tools and context providers. You still need agent logic, a model, and a controller.
- Do I still need guardrails with MCP?
- Yes. More tools mean more side effects to gate. MCP makes tool access easier; guardrails make it safe.
- Can I use MCP with any model?
- Yes. MCP is model-agnostic — it works with any model that supports tool calling.