The Words You Need to Talk About MCPs Without Pretending

My dad used to say you should know the rules, and then choose to break them. The classic example is who versus whom. To whom did you give the report? is technically correct. Saying it out loud will get you laughed out of most meetings. But if you don’t know the rule, you can’t make the choice.

MCPs are like that right now. The vocabulary is messy, half-formed, and inconsistently used. Half the people saying “MCP” mean three different things across three different sentences. You don’t need to correct them. But knowing the actual words, and using them precisely when it counts, is the difference between sounding like you know what you’re talking about and actually knowing what you’re talking about.

This piece is a working vocabulary for talking about MCPs in 2026. Not a glossary you memorize. A working vocabulary you absorb, the way a cook absorbs the names of cuts of meat. Once you have it, you stop fumbling, and you start having sharper conversations with vendors, engineers, and whoever is sitting next to you at the next AI panel.

For what it’s worth, this whole cluster was written for someone like me a few months ago: a BizOps practitioner who had been “using MCPs” for half a year and didn’t really understand what was happening under the hood. Attention’s own MCP server is one of the more capable in the conversation intelligence category, but the point of this piece is bigger than any vendor: the vocabulary problem belongs to the whole ecosystem, and the people who solve it for themselves first get to make better decisions.

TL;DR

• “MCP” gets used as shorthand for three different things: the protocol, the server, and the connection. The confusion is the source of half of all bad MCP conversations.
• The architecture has three roles: the host (your AI app), the client (the bridge), and the server (the vendor’s exposed tools). Vendors mostly ship servers. Anthropic mostly ships the standard and the reference clients.
• The capability surface has four primitives: tools, resources, prompts, and the manifest. Most vendors only meaningfully expose tools, but knowing what the others are makes you sharper at evaluating maturity.
• The cost vocabulary: context window, token, tool call, server-side processing. That vocabulary is what separates buyers who get burned from buyers who do not.
• The ecosystem vocabulary tells you where the standard is going. As of 2026, the answer is out of Anthropic’s hands and into a multi-vendor foundation.

Why does the MCP vocabulary feel so messy in 2026?

MCP went from “specification published by Anthropic” to “open standard stewarded by a foundation backed by every major AI lab” in fourteen months. [1] The vocabulary has not caught up. Different vendors use different words for the same thing. Marketing copy collapses precise terms into mush.

This is what happens with every new technical category. Cloud computing went through it. Containers went through it. Even “the internet” went through it: there was a long stretch where people used “web” and “internet” interchangeably and nobody minded much.

The difference now is the pace. MCPs went from idea to ubiquity in roughly the time it takes most teams to ship a single quarterly project. The vocabulary is being built in flight.

So here is the working vocabulary, organized by what you actually need to talk about. Six clusters, each unpacking the words you’ll hear in real conversations.

Cluster 1: “MCP” itself, and what people actually mean

MCP (Model Context Protocol). The open standard itself. Published by Anthropic in November 2024, donated to the Agentic AI Foundation in December 2025. [1] When someone says “MCP is an open standard,” they mean this.

MCP server. A specific implementation of the protocol that exposes a vendor’s tools to AI models. Attention has an MCP server. Canva has an MCP server. Notion has an MCP server. The phrase “our MCP” is technically meaningless, but everyone says it and everyone knows what they mean.

MCP connection. The configured link between a specific MCP server and a specific AI client (like Claude Desktop). An MCP connection can either be (for Claude) a connector or a custom setup: think of it like an app you download from the app store or one you have to get from the web.

The spec. The formal specification document at modelcontextprotocol.io, latest revision November 25, 2025. [2] The spec is the authoritative answer when vendors disagree about behavior.

The SDKs. Software development kits for building MCP servers and clients in different languages. [3] SDKs are not the protocol. They are conveniences that implement the protocol.

The grammar rule: the protocol is universal, the servers are vendor-specific, the connections are user-specific.

Cluster 2: The architecture words

Host. The application a person actually uses. Claude Desktop is a host. Claude.ai is a host. Cursor is a host.

Client. The bridge inside a host that speaks the MCP protocol to a server. Mostly invisible to users.

Server. The vendor’s deployed software that exposes their tools to MCP clients. Attention’s MCP server runs on Attention’s infrastructure.

The full chain: person → host → client → (over the wire) → server → vendor’s underlying APIs.

Transport. The wire protocol the client and server use to communicate. Three types in 2026: stdio (local), SSE (older remote), and Streamable HTTP (the production standard). [4]

JSON-RPC 2.0. The message format both sides use, regardless of transport. [5]

The grammar rule: hosts have clients, clients talk to servers, servers expose tools.

Cluster 3: The capability words

Tool. A specific function the AI can call through the MCP server: like search_calls or update_scorecard.

Resource. A piece of context the server can provide to the AI without being explicitly called. [6] Most vendors do not meaningfully expose resources yet.

Prompt. A pre-built prompt template the server provides for common workflows. [7] Even rarer in production than resources.

Manifest. The full inventory of tools the server exposes, returned when an AI client first connects. The manifest gets loaded into the AI’s context window on every request.

Capability. A formal declaration of what a server supports. Negotiated between client and server when the connection is established.

Sampling. A capability where the server asks the model to think on the server’s behalf. Rare in 2026 production servers.

The grammar rule: a server exposes tools, the manifest lists them, the AI reads the manifest, then calls tools.

Cluster 4: The cost words

Context window. The total amount of text an AI model can “see” at once. Claude Sonnet 4.6 has a 200,000-token context window. [9]

Token. The unit of text AI models use to count. Roughly 0.75 words per token in English.

Tool definition. The schema for a single tool. Each runs a few hundred to roughly a thousand tokens. A server with 100 tools loads a heavy manifest; a server with five loads a light one.

Tool call. A single invocation of a tool by the AI. The number of tool calls per real-world question is the single biggest cost variable in MCP usage.

Multi-call chain. When the AI strings together many tool calls to answer one question. Long chains are expensive, slow, and lossy.

Server-side processing. Where the server does the analytical work and returns a finished answer in a single tool call. Anthropic’s engineering team has documented cases where this reduces token usage by up to 98.7% compared to naive multi-call chains. [10]

Code execution (Code Mode). A pattern where the AI writes a short program that orchestrates multiple tool calls inside a sandbox, keeping intermediate results out of the model’s context window. [10]

The grammar rule: tokens are the unit, the context window is the container, tool calls are the verbs, and server-side processing is the escape hatch.

Cluster 5: The trust words

OAuth. The standard authentication framework MCP servers use to verify who is connecting.

Scope. What the OAuth token is allowed to do. Fine-grained scopes let you authorize specific capabilities rather than handing over the keys to the entire account.

Identity. The user (or service account) the MCP server is acting as. The question “whose identity is this running as?” is the foundation of every permission decision downstream.

Permission inheritance. Whether the MCP server respects the underlying product’s permission model. A well-designed server inherits permissions cleanly.

Personal access vs. shared access. Personal access means each user authenticates individually. Shared access means a single organization-wide token. Gong’s MCP server offers both. [12] Attention’s offers only personal access. [13]

The grammar rule: the question is never “is it secure” but whose identity is acting, with what scope, and does the permission model carry through.

Cluster 6: The ecosystem words

Registry. A public directory of available MCP servers. The largest community registry is mcp.so. Anthropic maintains its own curated directory of 75+ connectors. [1]

Connector. Anthropic’s product term for a curated MCP integration in the Claude app. Every connector is an MCP server, but not every MCP server is a connector.

Integration. The general-purpose word. An integration could be an MCP server, an API integration, a Zapier zap, or anything else that connects two systems.

AAIF (Agentic AI Foundation). The directed fund under the Linux Foundation that now stewards MCP. [1] Co-founded by Anthropic, Block, and OpenAI in December 2025.

Stewardship. The formal role of maintaining and evolving a standard on behalf of the community. MCP moved from one company’s spec to a multi-vendor foundation in late 2025.

Open standard. A specification that’s publicly documented, freely implementable, and maintained by a neutral body. MCP qualifies on all three counts.

The grammar rule: registries are how you find servers, connectors are how Anthropic curates them, AAIF stewards the standard.

A worked example: the vocabulary in action

Buyer: “So you have an MCP. What does it do?”

Seller: “We have an MCP server, yes. It exposes 68 tools across 15 functional groups. The most important is ask_attention, which does server-side processing for analytical questions. That keeps tool calls low and costs predictable.”

Buyer: “What’s the auth model?”

Seller: “OAuth, with personal access only. Every tool runs under the authenticated user’s identity, and permission inheritance is enforced.”

Buyer: “And the transport?”

Seller: “Streamable HTTP for production. SSE is deprecated.”

That conversation took 90 seconds and resolved more about the vendor’s actual capabilities than 30 minutes of marketing copy would have. The vocabulary did the work.

Alphabetical glossary

• AAIF (Agentic AI Foundation): directed fund under the Linux Foundation that stewards MCP
• Capability: formal declaration of what an MCP server supports
• Client: the protocol-speaking bridge inside a host
• Code execution (Code Mode): AI writes code to orchestrate tools inside a sandbox
• Connection: configured link between a server and a client
• Connector: Anthropic’s product term for a curated MCP integration
• Context window: total amount of text an AI can see at once
• Host: the application a person actually uses (Claude Desktop, Claude.ai, Cursor)
• Identity: user or service account the MCP server is acting as
• Integration: general-purpose word for any connection between two systems
• JSON-RPC 2.0: wire-format standard MCP uses for messages
• Manifest: full inventory of tools exposed by a server, loaded on every request
• MCP (Model Context Protocol): the open standard itself
• MCP server: a vendor’s implementation that exposes their tools
• Multi-call chain: AI strings together many tool calls to answer one question
• OAuth: standard authentication framework MCP servers use
• Open standard: specification publicly documented, freely implementable, neutrally maintained
• Permission inheritance: whether the MCP server respects the underlying product’s permission model
• Personal access: OAuth pattern where each user authenticates individually
• Prompt: pre-built prompt templates a server provides (rare in production)
• Registry: public directory of available MCP servers
• Resource: context a server provides without being explicitly called (rare)
• Sampling: server asks the model to think on its behalf (rare)
• Scope: what an OAuth token is allowed to do
• SDK: software development kits for building servers and clients
• Server-side processing: server does analytical work, returns finished answer
• Shared access: OAuth pattern where a single organization-wide token is used
• Spec: formal specification document at modelcontextprotocol.io
• SSE (server-sent events): older remote transport, replaced by streamable HTTP
• stdio: local transport for MCP servers on your own machine
• Stewardship: formal role of maintaining a standard on behalf of the community
• Streamable HTTP: newer remote transport, production default in 2026
• Token: unit of text AI models count and charge by
• Tool: specific function the AI can call through the MCP server
• Tool call: single invocation of a tool by the AI
• Tool definition: schema for a single tool, included in the manifest
• Transport: wire protocol clients and servers use to communicate

References

1. Anthropic — “Donating the Model Context Protocol and establishing the Agentic AI Foundation” — anthropic.com — 2025
2. Model Context Protocol — “Specification (2025-11-25)” — modelcontextprotocol.io — 2025
3. Model Context Protocol — “SDKs” — modelcontextprotocol.io — 2026
4. Model Context Protocol — “Transports” — modelcontextprotocol.io — 2025
5. JSON-RPC — “Specification 2.0” — jsonrpc.org — 2010
6. Model Context Protocol — “Resources” — modelcontextprotocol.io — 2025
7. Model Context Protocol — “Prompts” — modelcontextprotocol.io — 2025
8. Attention — “What Is an MCP Server? The Magic Problem, Explained.” — attention.com — 2026
9. Anthropic — “Claude Model Cards” — docs.claude.com — 2026
10. Anthropic Engineering — “Code execution with MCP” — anthropic.com — 2025
11. Attention — “The Three Ways an MCP Server Fails You” — attention.com — 2026
12. Gong Help Center — “About Gong MCP server” — help.gong.io — May 11, 2026
13. Attention Docs — “Attention MCP Server” — docs.attention.com — 2026
14. Attention — “Gong vs. Attention: A Side-by-Side Look at Their MCP Servers” — attention.com — 2026

Harry Abram owns inbound growth at Attention. Previously led ops at Nostra AI and was an early team member at Saturn (acquired by Snap). Georgetown SFS.