What is MCP and why does it matter for your work?

Definition box

Model Context Protocol (MCP) is an open standard that connects AI applications to external data and tools. It gives applications a common way to discover available capabilities, retrieve relevant information and request actions, subject to the connected systems' permissions and controls.

That is the useful forty-word definition. My judgement: MCP matters when copying material into a chat becomes the bottleneck, or when the task needs an action in another system. It is a connection standard, not a guarantee that the resulting work is correct.

The official introduction describes connecting AI applications with data, tools and workflows. Anthropic introduced MCP on 25 November 2024. Its origin does not make it exclusive to Claude.

Analogy

The MCP project uses the analogy of a USB-C port for AI applications. A shared connection convention makes it easier to connect different applications and systems without designing every pairing from scratch.

The analogy explains interoperability. It does not establish suitability for your work. Before connecting a system, identify what information should cross the boundary and which actions you actually want available.

A connector is the product-facing connection you configure. An MCP server is the software behind that connection, exposing capabilities to an AI application. MCP is the protocol they use to communicate. Vendors also use “connector” more broadly, so the label alone does not prove an integration uses MCP.

Red Hat's explainer describes the client/server approach and its value in reducing custom integrations. Think of the server as a service desk with a defined menu. Having access to the desk does not mean every operation in the underlying business system is available.

How it works (handshake/capability discovery, plain words)

The AI application hosts an MCP client. The client communicates with a server that exposes tools, resources or prompts. A local server can run as a process on your computer; a remote server is reached over a network. The architecture guide describes both arrangements and recommends OAuth for obtaining authentication tokens in the HTTP setting.

The three server capabilities have different jobs:

Capability	Plain meaning	Illustrative use
Tools	Operations the application can call	Search records or create a task
Resources	Information available as context	Read a document or a data schema
Prompts	Reusable interaction templates	Start a standard review request

“Handshake” needs a qualification. The older lifecycle specification describes an initialisation exchange. The current architecture page instead describes stateless requests carrying version and capability metadata, with a discovery request available to learn server capabilities. Calling every current interaction a mandatory initial handshake would be misleading.

In ordinary language, the application establishes what the server offers, lists relevant operations and requests one with the necessary inputs. The server returns a result. Authentication and product controls determine what access is available; discovery is not permission to do everything listed.

Imagine asking for an account briefing. The connection might retrieve relevant records, after which the assistant prepares a draft. My review would check which account was selected, whether the records are current and whether contradictory information was preserved. Those are workflow checks, not extra capabilities supplied by the protocol.

What you can do with it (5 role examples)

These examples describe possible workflows when suitable server tools and permissions exist. They are not promises that every product or connector supports every step.

A client adviser: retrieve approved account notes and recent tasks, then prepare a meeting brief with source references. I would keep the final interpretation and any client message with the adviser. The first useful trial is a single account whose history the reviewer already knows.

An operations manager: gather outstanding issues and draft a weekly exception report. Ask the assistant to distinguish missing information from a confirmed delay. Review the report before creating follow-up work, particularly where a ticket's owner or priority could affect another team.

A project lead: read the relevant calendar and project records to prepare an agenda. Separate factual scheduling information from suggested priorities. Let the lead decide what belongs in the meeting and whether any invitation should change.

A researcher: retrieve authorised documents and produce an evidence table. Require a source for each substantive claim and a separate column for uncertainty. Connection access helps gather material; it does not establish whether the source is reliable or whether the interpretation is sound.

A service team: retrieve a case and draft a handover or proposed next action. Check identity, case scope and unresolved questions before updating a shared record. Begin with preparation, then decide whether the observed quality justifies delegating an update.

This is the shift from Craft Assist, where AI helps me think and I decide, towards Augment, where AI carries several steps and I review. Adapt means AI works on files within rules I set while I stay on call. Adding a connection does not automatically move a workflow into Adapt.

Use the seven-layer map: model, harness, connections, skills, product, agency, my role. MCP belongs in connections. A skill can explain your process; the harness organises execution; the product determines available controls. My role remains explicit even when those layers work smoothly. See skills vs MCP vs plugins.

Support across tools (live table)

This is a dated snapshot for the connectors tracker, not an automatically updating feed. I started from the seed's Connections / MCP column and checked vendor documentation. A verified “yes” applies to the named surface, rather than every product carrying the same brand.

Tool	Tracker starting point	Official-page finding
Claude chat and Cowork	Connectors and plugins	MCP connections documented in Cowork overview
Claude Code	MCP not re-verified	MCP support confirmed in Code overview
Cursor	Enterprise MCP access controls	General MCP support confirmed in Cursor documentation; specific Enterprise controls not checked here Not yet confirmed
GitHub Copilot	Cloud agent MCP integration	GitHub docs confirm MCP across supported clients, including repository configuration for cloud agent
ChatGPT Work / ChatGPT	Apps and Developer Mode beta	Current help documents custom MCP connections for Business and Enterprise/Edu on web; the current Plugins flow does not require Developer Mode
Codex	MCP not yet confirmed	Official MCP documentation confirms local and remote server support on Codex clients
Gemini app	MCP not yet confirmed	Consumer app MCP support remains Not yet confirmed; Gemini API documentation confirms remote MCP for the Interactions API only

I have not converted every “connections” entry into MCP support. The seed also lists Microsoft 365 Copilot and Notion connections without establishing their exact MCP client scope. Those product-specific MCP claims remain Not yet confirmed. Perplexity Comet has no verified MCP entry in the seed Not yet confirmed.

Before copying someone else's setup, check the specific surface, server, authentication method and administrator policy. See what each tool can connect to for the practical next step and Skills vs MCP vs plugins for how the pieces fit together.

Risks in one paragraph

MCP is useful with appropriate controls, but it is not inherently safe. Stanford's security guide highlights hidden instructions in retrieved content, changed server behaviour, excessive privileges and credential risks. Prompt injection can turn apparently ordinary source material into attempted instructions. OAuth helps manage authorisation; it does not judge an action's business purpose. My starting rule is to grant only the access needed, keep consequential writes reviewable and name an owner who can disable the connection. Check the actual permissions and recheck them when the connection changes. Read the FAQ below ("Is MCP safe for work?") before extending a workflow into sensitive systems.

FAQ

What is an MCP server versus a connector?

An MCP server exposes capabilities; a connector is the connection a product lets you configure to reach them. MCP supplies the communication rules. The terms are related, but vendor naming varies. Check the underlying protocol and available operations before assuming a connector can use the same setup as another product.

Why does MCP matter if I can upload files?

MCP matters when your workflow needs access to a connected source or an action there, rather than a one-off copy. Uploading can be sufficient for a bounded review. Choose the simpler arrangement that meets the task, then judge whether a maintained connection improves evidence gathering enough to justify its access.

Does MCP let AI edit my calendar or CRM?

It can, when a server exposes the relevant operations and your product, account and permissions allow them. MCP itself does not promise a particular calendar or CRM action. Check the discovered tools and trial the workflow with harmless examples before delegating changes that affect colleagues, customers or shared records.

Which AI tools support MCP?

Claude, Claude Code, Cursor, GitHub Copilot, ChatGPT and Codex have documented support in the surfaces listed above. Google's API also documents remote MCP, but that does not verify consumer Gemini app support. Use the dated table as a starting point and confirm the exact client before installing a server.

Is MCP safe for work?

It can be appropriate for work when the server, permissions and actions fit your organisation's rules. Treat retrieved content as potentially untrusted, inspect the access request and retain a review point for consequential changes. A familiar vendor name or standard protocol does not remove the need to check the actual workflow.

Connections are a Craft layer: learn what to connect and what to keep human.

Verification notes

• Gemini consumer-app MCP support is not yet confirmed. API support is a separate, verified finding.
• Cursor's precise Enterprise access controls, Microsoft 365 Copilot and Notion product-specific MCP client scope, and Comet MCP support remain not yet confirmed here.
• The seed's older ChatGPT Developer Mode wording is superseded for the current flow described in the opened help page. No individual-plan custom-server entitlement is inferred.
• The current MCP architecture and an older lifecycle specification use different connection flows. The article distinguishes them rather than presenting the older handshake as universal.
• Role examples and Craft guidance are editorial proposals, conditional on tools and permissions.