Short answer
The Model Context Protocol (MCP) is an open standard for connecting AI applications to external tools and context through tools, resources, and prompts. In its host-client-server model, the host application creates client instances that communicate with MCP servers exposing those capabilities. Anthropic introduced MCP in November 2024 and donated it in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation. MCP has broad implementation support and a large public server ecosystem.
MCP standardizes capability access, not agent execution or agent-to-agent messaging. A2A addresses communication between independent agents. UHP is a younger, open, versioned standard for the configured-harness boundary: it defines how a product sets up, selects, and drives a harness, including its model, skills, MCP servers, and which tools are available, as well as tasks, sessions, streaming, files, and cancellation. These layers compose: a product uses UHP to set which MCP servers the harness has and then to drive it, while MCP carries runtime calls to tools exposed by those servers.
- MCP is an open standard for connecting an agent to tools, resources, and prompts; it operates inside the agent's loop.
- It uses a host-client-server model: a host creates client instances that communicate with MCP servers exposing capabilities.
- MCP is an established, widely adopted standard, now hosted by the Agentic AI Foundation under the Linux Foundation.
- MCP is one layer among several: it composes with A2A (agent-to-agent) and UHP, which standardizes harness configuration and product-to-harness driving, rather than replacing either.
MCP in one sentence
The Model Context Protocol is an open standard that gives an AI agent a uniform way to reach tools, data, and reusable prompts, so a capability built once can be used by any MCP-aware application. It is often described by analogy to a universal port: instead of a bespoke integration per tool, a model speaks one contract to many capabilities.
What the Model Context Protocol standardizes
MCP standardizes three primitives an agent consumes: tools it can call, resources it can read, and prompts it can reuse. It defines how a client discovers what a server offers and how calls and results are exchanged. It deliberately does not define how a configured harness is assembled, selected, or driven as a whole; UHP covers that separate boundary.
MCP architecture: host, client, and server
MCP uses a host-client-server model. The host is the AI application or agent harness and creates client instances that communicate with MCP servers exposing tools, resources, or prompts. In the 2026-07-28 specification, the protocol core is stateless: the initialization handshake and protocol-level sessions were removed, requests are self-contained, and capabilities are declared per request. Optional extensions require explicit support from both sides; MCP Apps and Tasks sit in that extension layer.
- Host
- The AI application or agent harness the user interacts with.
- Client
- An MCP client created by the host to communicate with a server; under the 2026-07-28 protocol, HTTP requests do not require a protocol-level session.
- Server
- A process exposing tools, resources, or prompts to the client.
Tools, resources, prompts, and extensions
- Tools: callable actions the agent can invoke, such as querying a database or calling an API.
- Resources: readable data the agent can pull into context, such as files or records.
- Prompts: reusable prompt templates a server can offer to the client.
- Extensions: optional capabilities such as UI surfaces (MCP Apps) and long-running tasks (Tasks), which require explicit support from both sides.
How an MCP tool call works
At a high level: the client connects to a server and discovers its tools; the agent decides to call one; the client sends the call to the server; the server executes and returns a result; the client places that result back into the agent's context. The contract is uniform, so the same client can talk to any conformant server without custom glue.
Security and trust boundaries
Because MCP servers can execute actions and expose data, each server is a trust boundary. An MCP client should treat servers as it would any external dependency: least privilege, explicit approval for sensitive actions, and care with untrusted servers, whose tool descriptions and outputs can attempt to steer the agent. MCP defines the contract; the surrounding application still owns authorization and isolation.
What MCP does not standardize
- How an agent reasons or plans.
- Communication between independent agents.
- How a product configures, selects, invokes, and controls a complete agent harness.
- Frontend rendering and user interaction.
When to use MCP
Reach for MCP when you need runtime access to tools and data through a standard contract, so capabilities are reusable across applications. If your question is how a product sets up, selects, and drives a harness, including its model, skills, MCP servers, and available tools, that is UHP's configured-harness boundary; see UHP vs MCP.
Implementation checklist
- Decide which capabilities become tools, which become resources, and which become prompts.
- Treat every server as a trust boundary with least-privilege access.
- Send the protocol version and per-request capabilities with each 2026-07-28 request, and advertise optional extensions explicitly.
- Implement protocol authorization where applicable, and enforce consent, least privilege, and server isolation in the surrounding application.
FAQ
- What does MCP stand for in AI?
- MCP stands for Model Context Protocol. It is an open standard for connecting an AI agent to tools, resources, and prompts through a uniform client-server contract, so a capability built once can be reused across MCP-aware applications.
- Is MCP a protocol, framework, or API?
- MCP is a protocol: an open standard that defines how a client and server exchange tools, resources, and prompts. It is not a framework you build inside, and it is more structured than a single REST API, since it defines discovery and a uniform contract across many servers.
- Does MCP run an AI agent?
- No. MCP connects an agent to capabilities; it does not run the agent. The harness owns the loop. UHP standardizes how a product configures and drives that harness, while MCP remains the runtime wire for tools exposed by configured MCP servers.
- What is the difference between MCP and a REST API?
- A REST API is a bespoke interface to one service. MCP is a uniform contract across many tool servers, with built-in discovery, so an MCP client can talk to any conformant server without custom integration per service.
- How do MCP and A2A work together?
- They govern different boundaries. MCP connects an agent to tools inside its own loop; A2A connects independent agents to each other. A system can use both: an agent reaches tools over MCP while communicating with other agents over A2A.
- What is an MCP host?
- An MCP host is the AI application that coordinates MCP clients, applies consent and security policies, and integrates server-provided tools, resources, and prompts into the model workflow.
- Is MCP controlled by Anthropic?
- MCP was introduced by Anthropic in 2024 and donated in 2025 to the Agentic AI Foundation under the Linux Foundation, so it is now governed as an open standard within that foundation rather than by a single company.
See how a product configures and runs a whole harness
Through UHP, a product configures and drives a harness; MCP carries runtime calls to tools exposed by the MCP servers configured on that harness.
Read the agent harness guide