Automation10 min
What Is MCP? Model Context Protocol for Marketers
Owen Nash · Marketing Automation Expert
MCP explained for marketers: what the Model Context Protocol does, why AI agents need it, and official registry data we pulled on 10 Sep 2026.
What MCP Is, and Why Marketing Agents Need It
MCP — the Model Context Protocol — is an open standard for connecting AI applications to external systems: data sources, tools, and workflows. The protocol’s own documentation uses a deliberately physical analogy: MCP is “like a USB-C port for AI applications,” giving models one standard way to plug into the systems where your work already lives (modelcontextprotocol.io, fetched 10 September 2026). For a marketing team, that is the difference between an assistant that can only write and one that can read your CRM, pull last week’s ad numbers, or draft a reply from a real support thread.
Anthropic open-sourced the protocol on 25 November 2024, describing it as “a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments” (anthropic.com). Adoption has spread well beyond Anthropic: the MCP documentation now lists Claude and ChatGPT among supported assistants and Visual Studio Code and Cursor among supported developer tools, and the current specification version is 2026-07-28.
The problem MCP removes is arithmetic. If every AI tool needs a bespoke connector to every system, cost scales as tools × systems — and most marketing teams run a dozen systems before they count ad platforms. A shared protocol collapses that into one integration per system, which is why “does it speak MCP?” is turning into a line item in software buying rather than a developer-only concern.
Agentic marketing — hand a task to an AI agent, get a finished outcome — usually fails at the same place: access, not intelligence. The models are good enough; they simply cannot see the systems that hold the work. HubSpot’s 2026 marketing statistics page, which draws on a State of Marketing survey of 3,400 marketers globally, puts the gap in numbers: 80% of marketers already use AI for content creation and 75% use it for media production, yet only 47.18% strongly or somewhat agree they understand how to incorporate AI into their marketing strategy, and just 47.63% say the same about measuring its impact (hubspot.com/marketing-statistics, fetched 10 September 2026). Near-universal adoption, half the teams unsure how to make it pay.
A second figure raises the stakes. The same page reports that nearly 70% of marketers see leads arriving later in the buying process because buyers do more AI-assisted research first. When the prospect shows up already informed, the team that answers fastest wins — and speed comes from an agent that can pull the account history, check stock, and draft a reply without a human copying data between five browser tabs.
Before MCP, wiring an agent to even one SaaS system meant a custom integration, a fresh security review, and a maintenance burden every time the vendor changed an API. That economics is why so many “AI marketing agents” stayed in demo videos: connecting was the expensive part, and it had nothing to do with how smart the model was.
The Three Building Blocks, Translated for Marketers
An MCP server exposes its capabilities through three building blocks — Tools, Resources, and Prompts — and the documentation assigns each a different controller: the model, the application, and the user (server concepts, fetched 10 September 2026). That assignment is not academic. It tells you precisely which surfaces need a human gate.
- Tools are actions the model can decide to call on its own judgment — “create a HubSpot contact from this form fill”, “send the weekly summary to Slack”, “schedule this post for Thursday”. Tools can write, so these are the calls you gate and log.
- Resources are passive, read-only context the application loads: your brand guidelines, a keyword list, the current campaign brief. Exposing them broadly is usually fine; the risk is stale data, not destroyed data.
- Prompts are user-invoked templates — “draft the monthly performance review from these numbers”. Because a person starts them, prompts are the right place to standardize how your team asks for recurring work.
- Transports decide who can use the server. Local servers speak STDIO and typically serve one client on one machine; remote servers speak Streamable HTTP and can serve many clients (architecture). That is the line between a personal desktop helper and a shared team capability.
What the Official MCP Registry Shows Today
To check how much of this is real rather than promised, we queried the official MCP Registry directly on 10 September 2026 through its public API (registry.modelcontextprotocol.io, the /v0/servers endpoint). Method note: we paginated the API’s default listing and read the entries returned. The registry is a public directory, not a curated catalogue, and it changes daily — treat every number below as a point-in-time snapshot.
- Scale: the default listing returned more than 29,000 entries before we stopped paginating (that count includes multiple versions of the same server, so it overstates the number of distinct servers), which says the ecosystem is far past hand-curated lists.
- CRMs are covered, unevenly. A search for “hubspot” returned 13 listings across four distinct connectors, including one described as covering “HubSpot CRM: contacts, deals, tickets, line items, properties, marketing, files, workflows, threads.” HubSpot does not publish that connector — a community developer does, which changes how you assess the risk.
- First-party servers are arriving. A search for “notion” returned an entry labelled simply “Official Notion MCP server” (com.notion/mcp) — evidence that vendors are beginning to ship connectors themselves rather than leaving it to volunteers.
- The long tail is already crowded. Searches for “seo”, “social”, “email”, “analytics”, and “crm” each returned 70–100+ listings on the same day, spanning narrow tools from SEO audits to “SEO + GEO” visibility checkers. Crowded, however, is not the same as trustworthy.
- The honest caveat: a registry listing is not a security review, an SLA, or a support contract. Most entries are individual projects. Treat a listing as a lead to investigate, not a badge you can rely on.
Step by Step: Connect One Server, Then Gate the Writes
The fastest way to understand MCP is to connect one low-risk server and watch a tool call happen. This walkthrough follows the setup in the protocol’s own documentation and uses Claude Desktop, although the pattern is the same in any MCP-capable client. Steps one to five are safe to run today; steps six and seven are the gate before anything gets write access.
- Step 1 — Pick a read-only job. Something like “summarize last month’s campaign briefs”. Read-only output keeps the blast radius small while you learn what the tool calls actually do.
- Step 2 — Choose the server and check who built it. Look at the registry entry, the linked repository, and when it was last updated. Prefer an official server (the Notion example above) or a well-maintained project over an unlisted one.
- Step 3 — Add the configuration. In the client’s config file you declare an mcpServers block; the documentation’s filesystem example points the server at specific folders — {"mcpServers":{"filesystem":{"command":"npx","args":["-y","@modelcontextprotocol/server-filesystem","C:\\Users\\you\\Projects"]}}}. Scope it to one folder, never an entire drive.
- Step 4 — Restart the client and confirm the tools appear. The documentation is explicit that the application must fully restart to load the server; you should see the server’s tools listed before you invoke anything.
- Step 5 — Run the job, then read the tool-call log. Every action the agent takes should be inspectable. If it does something you did not expect, stop and narrow the server’s scope before continuing.
- Step 6 — Scope the credential and know the attack classes. Give each server its own least-privilege token. The protocol’s Security Best Practices document catalogues exactly what this defends against — the Confused Deputy Problem, Token Passthrough, Server-Side Request Forgery, State Handle Hijacking, and Local MCP Server Compromise (MCP security, fetched 10 September 2026). None of these are exotic; they are the standard reasons to slow down on a new vendor integration.
- Step 7 — Require human approval for writes, and hand deterministic steps to your workflow tool. Reading can be liberal, but creating a contact, emailing a prospect, or moving budget should pause for a click. Teams already running n8n or Zapier typically use MCP for the judgment step and let the workflow tool handle logging, routing, and notifications — our n8n marketing workflow blueprints are a reasonable place to attach that hand-off.
Worked Example: One Brand, One Read-Only Server
Scenario (illustrative example): Marlow & Finch, the fictional DTC skincare brand we use across these guides, runs a small team on HubSpot for CRM, Google Sheets for reporting, and Slack for internal comms. Their first MCP task is deliberately unglamorous — every Monday morning, summarise the previous week’s campaign briefs and flag anything that changed.
Rather than a developer integration, the team connects one read-only server: a filesystem server scoped to a single folder called /campaign-briefs, attached to an MCP-capable assistant. The agent lists the folder (a Resource), reads the three briefs, and returns a summary with the changes it spotted. Nothing is written anywhere, no customer record is touched, and the whole setup takes an afternoon.
The payoff is not the summary — it is what the team learns from the tool calls. They can see exactly which files the agent opened, how it interpreted “changed”, and where it had to ask a clarifying question. That visibility is the real output of a first MCP project. When the team later adds a server that can create HubSpot tasks, they already know how the agent behaves and where the approval gate belongs — which is the difference between an agent you trust and an agent you babysit.
FAQ: MCP for Marketers
Q: What is MCP in one sentence? A: The Model Context Protocol is an open standard that lets AI applications connect to external systems — files, databases, SaaS tools, and workflows — through one shared interface, so an AI agent can read your data and take actions instead of only producing text.
Q: Do I need a developer to use MCP? A: To consume an existing server, usually no — adding a remote or local server is configuration, not coding. To build a private server for an internal system, yes. Start by consuming a read-only server before you invest in building anything.
Q: Is it safe to connect an MCP server to my CRM? A: Treat it like any vendor integration, with one extra rule. Give it a scoped, least-privilege credential; start read-only; require human approval for writes; and prefer an official or well-maintained server. The protocol’s own security document lists the specific attack classes to check for, and they are the standard reasons to move carefully.
Q: How is MCP different from Zapier or n8n? A: Workflow tools connect apps through triggers and actions you design in advance. MCP gives a model the ability to choose and call tools at runtime. They are complementary: many teams keep Zapier or n8n for deterministic steps and use MCP for the steps that need judgment.
Q: Which clients support MCP today? A: The protocol documentation lists assistants including Claude and ChatGPT and developer tools including Visual Studio Code and Cursor. Support is spreading quickly, which is exactly why the integration is worth learning once — see our roundup of AI marketing agents for where this is heading.
What to Do This Week
Three actions, in order. First, pick one read-only job in your stack and connect a single server to it — the tool-call log you get back teaches you more than any explainer. Second, audit who can act on your behalf: list every agent or integration with write access to a customer system, and confirm each one holds a scoped credential rather than a shared key. Third, add one question to your next tool evaluation — “does it speak MCP, and who publishes the server?” The protocol is not a feature you admire from a distance; it is the connector layer your next agent platform will be judged on.
Key Takeaways
- 1MCP is an open standard — a “USB-C port for AI” — that lets agents read your systems and act, not just write.
- 2HubSpot 2026: 80% of marketers use AI for content and 75% for media, yet only about 47% can measure its impact.
- 3Our 10 Sep 2026 registry check: 29,000+ listings; “hubspot” returned four distinct connectors, “notion” an official one.
- 4Start read-only: add one server, watch the tool-call log, then scope a least-privilege credential before granting write access.
- 5MCP and workflow tools like n8n or Zapier are complements — MCP handles judgment, workflows handle deterministic steps.
Tags:Model Context ProtocolMCP for marketersAI marketing agentsMCP servers
Owen Nash
·Marketing Automation ExpertOwen Nash is a marketing automation specialist with deep expertise in AI agent workflows. He has built automation systems for companies ranging from startups to Fortune 500 enterprises.
Recommended AI Tools
Zapier
Free & PaidNo-code automation platform connecting AI tools with your marketing stack.
Automation
ChatGPT
Free & PaidAI-powered conversational assistant for content creation, research, and marketing copy.
Content
Claude
Free & PaidAnthropic's advanced AI assistant for analysis, writing, and complex marketing tasks.
Content