Host
The AI application that coordinates and manages one or more MCP clients, for example a coding assistant or a chat application.
MCP architecture overviewThe open standard that connects AI applications to tools and data: hosts, clients, servers, the current specification and its security implications.
The Model Context Protocol (MCP) is an open-source standard for connecting AI applications to external systems such as data sources, tools and workflows.
An AI application, the host, runs one client per MCP server, and the server offers tools, resources and prompts over JSON‑RPC 2.0. MCP standardizes the connection and leaves enforcement of consent and safety to implementers.
The AI application that coordinates and manages one or more MCP clients, for example a coding assistant or a chat application.
MCP architecture overviewA component inside the host that maintains a connection to one MCP server and obtains context from it for the host to use.
A program that provides context to clients. It can run locally over the stdio transport, or remotely over Streamable HTTP, where it typically serves many clients.
Servers expose tools (executable functions), resources (data for context) and prompts (reusable templates). Clients can offer elicitation, which lets a server request input from the user.
Revision 2026‑07‑28 changed how every request works, which matters to anyone logging or filtering MCP traffic.
The initialize handshake and protocol-level sessions are gone. Every request now carries its protocol version and the client’s capabilities.
Key changes in this revisionServers must implement server/discover to advertise their supported protocol versions, capabilities and identity.
Roots, Sampling and Logging are deprecated. So is OAuth Dynamic Client Registration, in favor of Client ID Metadata Documents.
The specification says descriptions of tool behavior should be considered untrusted unless they come from a trusted server. The model reads those descriptions when choosing what to call.
MCP specificationHosts must obtain explicit user consent before invoking any tool, and the specification acknowledges that MCP itself cannot enforce its security principles at the protocol level.
A local MCP server is a program executed on the user’s machine. MCP’s security guidance warns of malicious startup commands and recommends sandboxing and showing the exact command before it runs.
MCP security best practicesThe guidance forbids token passthrough, where a server accepts tokens that were not issued for it, and recommends minimal scopes that are elevated only when needed.
Straight answers, kept current. Last updated .
Ask us anything →An open-source project with its own governance. Its governance page states that the project is established as Model Context Protocol a Series of LF Projects, LLC. Anthropic donated MCP in December 2025 to the Agentic AI Foundation, a directed fund under the Linux Foundation. Lead Maintainers and Core Maintainers steer the specification, and changes are proposed as Specification Enhancement Proposals (SEPs).
No, and the specification says as much. It states that MCP enables arbitrary data access and code execution paths, and that the protocol cannot enforce its security principles itself. User consent, access control, authorization and data protection are the job of the hosts, clients and servers that implement it. That is why an inventory of the MCP servers in use is the first control to put in place.
An API is an interface a developer codes against. An MCP server wraps capabilities, often existing APIs, in a standard form an AI application can discover: it lists its tools, resources and prompts, describes them in text and schemas, and executes the tool calls the model requests. The security-relevant difference is the caller: a model choosing actions from descriptions, not a developer writing fixed calls.
In part, and it is worth being exact. Cranium AgentSensor inventories agent frameworks, MCP servers and tools from source code into the AI‑BOM, so you know which servers your applications reference. The Tool Inspector shows the tool calls an agent made. Cranium AI does not run an MCP gateway or proxy and does not authorize or block tool calls at runtime. That control belongs at the host or the tool layer.
Primary sources for the figures and frameworks cited on this page. Last verified .
See how Cranium AI discovers, observes, governs, secures and proves every AI model, agent and vendor in your enterprise.
Cookie Policy · Privacy Policy
Global Privacy Control detected. Opt-out request honored: marketing tags stay off in this browser.