How B2B Marketers Use MCP to Run Their Whole Stack

Ask Claude which Google Ads campaigns produced closed-won pipeline last quarter. With MCP installed, Claude pulls deal data from HubSpot, joins it to spend data from Google Ads, and posts a ranked list to Slack from one chat window.

This guide explains MCP for B2B marketing and RevOps teams. No code background needed. By the end you’ll know what MCP servers do, which ones to plug into your stack first, how to set them up, how to build a custom one, and where to host it.

TL;DR

MCP, short for Model Context Protocol, is an open standard from Anthropic released in November 2024. MCP gives AI assistants a single way to connect to outside tools like HubSpot, Salesforce, LinkedIn Ads, Apollo, and GA4. B2B marketers use MCP for pipeline attribution, ABM campaigns, lead enrichment, and SDR sequencing. With six servers wired up, one prompt now runs an entire ABM launch end to end. Setup takes minutes per tool, no engineering team required.

Key Takeaways

MCP is a single protocol every AI tool agrees to speak, ending the need for one-off integrations
Anthropic released MCP in November 2024, and over 10,000 MCP servers existed by early 2026
B2B marketers run MCP for HubSpot, Salesforce, LinkedIn Ads, Google Ads, Apollo, ZoomInfo, Outreach, and GA4
Pipeline attribution, ABM, lead enrichment, and SDR sequencing all run through one chat now
Stitching MCPs together uses three patterns: sequential, parallel, and conditional chaining
SEO and AEO workflows for B2B intent queries run on Ahrefs, Semrush, Search Console, and Frase MCPs
One prompt now runs an end-to-end ABM campaign across six MCP servers with human approval gates
Building a custom MCP server takes 30 minutes with the official Python or TypeScript SDK
Security risks like tool poisoning and prompt injection are real, so connect trusted servers only

What is MCP in plain English?

MCP is a single set of rules every modern AI tool follows when talking to outside apps. The full name is Model Context Protocol. Anthropic released it in November 2024. OpenAI and Google adopted it within months.

Without MCP, every connection between an AI tool and an outside app needed custom code. Claude to HubSpot was one project. ChatGPT to HubSpot was a separate project. Repeat for every tool, every AI client. The result was thousands of one-off integrations, most of them broken or out of date.

With MCP, the integration is written once and works for every AI tool following the standard. HubSpot ships one MCP server. Claude, ChatGPT, Cursor, and every other MCP-compatible app reads from it.

For a B2B marketer, this means your AI assistant reads live data from your CRM, ad platforms, and sales engagement tools through one chat. No CSV exports. No tab switching.

How does an MCP server work?

An MCP setup has three pieces: the host, the client, and the server.

The host is the app you chat with. Examples: Claude Desktop, the Claude website, Cursor, ChatGPT.

The client is a small piece of code inside the host. It handles the back-and-forth between the AI and the outside tool.

The server is the connector to one specific outside tool. There’s an MCP server for HubSpot, one for Salesforce, one for LinkedIn Ads, and so on.

The flow when you ask a question:

You type a question in Claude
Claude checks which MCP servers are connected
Claude picks the right server and sends a request
The server fetches the data or runs the action
Claude reads the result and writes the answer back

Real example. You type, “Pull last quarter’s pipeline by source from HubSpot.” Claude finds the HubSpot MCP server, sends a structured request, gets the data back, and writes a summary in chat. Total time: 8 seconds.

Action item: Open Claude Desktop, look at the bottom of the chat box for the connectors menu, and count how many tools are connected today. This number is your starting baseline.

How does MCP differ from a regular API?

APIs are old. Most marketing tools have one. APIs let two programs talk to each other, but only after a developer writes code spelling out every step.

MCP sits on top of the API and adds a description the AI reads on its own. The description tells the AI what each tool does, what inputs it needs, and what comes back. The AI uses this to pick the right tool without help.

Practical difference for B2B marketers. With an API, your engineer writes a script. With MCP, you type a request in plain English and the AI figures out which tool to call.

MCP also keeps context across the chat. Run three reports in a row and the AI remembers the earlier numbers. APIs forget everything between calls.

Why should B2B marketers care about MCP?

B2B marketing has a data problem unique to the category. Six-month sales cycles. Multiple decision-makers per account. Data scattered across CRM, sales engagement, ad platforms, and intent providers. Most attribution questions take days to answer.

MCP collapses the lookup. Your AI reads live data straight from HubSpot, Salesforce, LinkedIn Ads, and Apollo at the same time. No CSV exports. No tab switching. No “I’ll get back to you on which channel sourced this deal.”

Real B2B workflows get faster. Pipeline attribution by source drops from 2 hours to 2 minutes. SDR enrichment for 50 net-new contacts runs in one chat instead of 3 tabs. ABM account research formerly taking a half day now runs while you grab coffee.

MCP-connected AI also takes action. Enrich a contact, build a target list, draft a sequence, post to Slack. Each of those used to need a different tab and a human to drive it.

Which MCP servers should B2B marketers install first?

Start with the eight servers covering most B2B marketing workflows. Each one targets a tool your RevOps stack already pays for.

The B2B starter stack:

HubSpot MCP for CRM data, contacts, deals, pipeline reports, and lifecycle stages
Salesforce MCP via Agentforce 3 for SOQL queries and CRM data with Einstein Trust Layer
LinkedIn Ads MCP for ABM display, sponsored content, and lead gen form data
Apollo MCP for prospecting, contact enrichment, and outbound sequences
ZoomInfo MCP for enterprise B2B data, intent signals, and account research
Google Analytics 4 MCP for site behavior, conversion paths, and source attribution
Google Search Console MCP for B2B intent keywords and ranking positions
Slack MCP for posting pipeline alerts and notifying SDR teams

Add Outreach or Salesforge MCP if you run sales engagement, Notion MCP for internal docs, and Zapier MCP as a fallback bridging 8,000+ apps.

As of March 2026, the major ad platforms with MCP servers include Google Ads, LinkedIn Ads, Meta Ads, Amazon Ads, and Microsoft Bing Ads. Apollo launched its official MCP server in February 2026. ZoomInfo exposes enterprise MCP access through its GTM Context Graph.

Action item: List the five B2B tools your team opens most days. Check whether each ships an official MCP server or has a community one. Bookmark the docs.

What does the HubSpot MCP server do?

The HubSpot MCP server gives Claude read and write access to your CRM. CRM objects accessible include contacts, companies, deals, tickets, carts, products, orders, line items, invoices, quotes, subscriptions, and segments. Engagements include calls, emails, meetings, notes, and tasks. Marketing and content includes campaigns and campaign metrics, landing pages, website pages, and blog posts.

Sample B2B prompts your RevOps team runs daily:

“Show me marketing-sourced pipeline by campaign for the last 90 days”
“What is the MQL-to-SQL conversion rate from Google Ads versus LinkedIn this quarter”
“Find all enterprise deals in negotiation stage older than 30 days with no activity in 14 days”
“Pull every deal closed-won in Q1 and group by original source plus first-touch campaign”
“Score every contact created last week by ICP fit using firmographic fields”

Setup uses OAuth in Claude. Open Settings, click Connectors, search HubSpot, log in. The HubSpot admin grants access first, then anyone on the account connects.

What does the LinkedIn Ads MCP server do?

The LinkedIn Ads MCP server connects ABM and demand gen campaigns to Claude. Multiple options exist: CData ships a managed server, Radiate offers a hosted MCP for read-only analysis, and GrowthSpree maintains a community version.

LinkedIn is the dominant paid channel for B2B, so MCP access matters. Most servers today offer read-only analysis with write operations rolling out. Use cases your team runs weekly:

“Compare CPA by job title across our Sponsored Content campaigns this quarter”
“Find audience segments with under 1% CTR and flag for creative refresh”
“Pull lead gen form submissions from the past 30 days and match to HubSpot contacts”
“Show ABM display performance against our top 50 target accounts”
“Diagnose why our VP-level campaign CPA tripled last week”

Combined with HubSpot MCP, you get cross-platform attribution in one chat. Ask Claude, “Which LinkedIn Ads campaign produced the most pipeline value last quarter,” and it joins ad spend to deal data automatically.

What does the Apollo MCP server do?

Apollo launched its official MCP server in February 2026 as a way to run outbound workflows entirely inside Claude. The server exposes tools for prospecting, contact enrichment, sequence creation, and CRM checks.

The full B2B sales workflow now runs in one chat. Define your ICP in plain English. Pull matching contacts. Enrich with verified emails and phone numbers. Check HubSpot or Salesforce for duplicates. Generate personalized outreach. Stage a sequence. Wait for human approval before launch.

Sample prompts:

“Find VPs of Sales at B2B SaaS companies with 50 to 500 employees in the United States”
“Enrich these 100 contacts with verified emails, direct dials, and recent job changes”
“Check which of these prospects already exist in HubSpot and exclude them”
“Create a 4-step sequence for VPs of Marketing at fintech companies, focused on attribution pain”

ZoomInfo’s MCP server exposes six tools: find accounts, enrich accounts, research accounts, find contacts, enrich contacts, and research contacts. Setup is a one-time server configuration using existing ZoomInfo API credentials.

What does the Google Analytics 4 MCP server do?

The GA4 MCP server lets Claude query 200-plus dimensions and metrics in plain English. The Google Analytics MCP Server connects Google Analytics 4 data to MCP clients like Claude and Cursor, allowing users to query their website traffic, user behavior, and other analytics data using natural language. By leveraging over 200 GA4 dimensions and metrics, the server translates natural language queries into API calls, retrieving and presenting data in a user-friendly format.

Sample B2B prompts:

“Which landing pages drive the highest demo request conversion rate”
“Compare account-level engagement on the pricing page over the past 90 days”
“Show me which pillar blog posts produced the most downstream MQLs”
“Which traffic sources have the longest session duration for enterprise visitors”

Two install paths exist. Google ships a CLI-based install for technical users. Coupler.io offers a browser-based OAuth setup connecting 370-plus sources, including GA4, alongside ad platforms in one MCP server.

How do you install an MCP server in Claude?

Claude offers two install paths: connectors for hosted servers and local config for custom ones. Pick the path matching the server type.

For hosted servers like HubSpot, Klaviyo, and Notion, use connectors. Open Claude.ai or Claude Desktop, go to Settings, click Connectors, find the tool, click Connect, and complete OAuth. Total time: under two minutes per tool.

For local servers, edit a JSON config file. Claude Desktop stores MCP configuration in a JSON file. The location depends on your operating system. macOS uses ~/Library/Application Support/Claude/claude_desktop_config.json. Windows uses %APPDATA%\Claude\claude_desktop_config.json.

Quickest way to find the file: open Claude Desktop, click Settings, choose Developer, and click Edit Config. The file opens in your default editor.

A working config for the filesystem server looks like this:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/yourname/Desktop"
      ]
    }
  }
}

Save the file, restart Claude Desktop, and look for the connector icon in the chat input. Click the icon to see active tools.

Action item: Install one connector this week using the Settings menu. HubSpot or Klaviyo if you use them, otherwise Notion or Slack.

Which B2B marketing workflows run on MCP today?

Eight workflows are common across B2B teams running MCP in production. Each replaces a multi-hour manual process with a single chat prompt.

Pipeline attribution by source. Connect HubSpot MCP plus Google Ads MCP plus LinkedIn Ads MCP. Ask: “Which paid channel produced the most closed-won pipeline last quarter, broken out by campaign and original source.” Claude joins three servers and returns one ranked table.

Target account research. Connect ZoomInfo or Apollo MCP plus HubSpot MCP. Ask: “Pull our top 50 target accounts from HubSpot. For each, find recent funding news, headcount changes, and the right VP-level contact to reach out to.”

ICP-based prospect lists. Connect Apollo MCP. Ask: “Find 200 RevOps Directors at B2B SaaS companies between 100 and 500 employees in North America who have raised funding in the past 12 months. Exclude anyone in HubSpot already.”

SDR sequence drafting. Connect Outreach or Salesforge MCP plus HubSpot MCP. Ask: “Build a 5-step sequence for VPs of Marketing at fintech companies. Pull our three best-performing past sequences for context. Personalize using firmographic and tech stack data.”

Lead scoring audits. Connect HubSpot MCP. Ask: “Show me MQLs from the past 90 days never converting to SQL. Identify the firmographic patterns most predictive of conversion failure.”

Cross-channel ABM diagnostics. Connect LinkedIn Ads MCP plus 6sense or Demandbase MCP plus GA4 MCP. Ask: “For our top 50 target accounts, show LinkedIn ad engagement, intent surge data, and on-site behavior. Flag accounts hot enough for SDR outreach this week.”

Lead enrichment at scale. Connect HubSpot MCP plus Apollo MCP plus a web search MCP. Ask: “Pull the 50 newest contacts from yesterday. Enrich with verified emails, direct dials, LinkedIn URLs, and recent company news. Update the CRM fields.”

Pipeline gap analysis. Connect HubSpot or Salesforce MCP plus Slack MCP. Ask: “Compare actual Q1 pipeline against our quota target. Identify the channels and segments under-pacing. Post the summary and three recommendations to #revops.”

Action item: Pick one workflow above. Write the prompt in plain English. Run the prompt this week and time how long the result takes versus your current manual process.

How do you stitch MCP servers together?

The real value of MCP shows up when you chain 3 or 4 servers in one prompt. The AI passes data from one server to the next, joins the results, and writes the answer.

Three chaining patterns cover most B2B work: sequential, parallel, and conditional. Each one has a different shape and a different use case.

Sequential chaining

The AI calls servers in order. The output of step 1 becomes the input of step 2.

Example: enrich and route a new MQL.

HubSpot MCP returns the contact created in the last hour
Apollo MCP enriches the contact with verified email, direct dial, LinkedIn URL, and recent job changes
ZoomInfo MCP adds firmographics (revenue, headcount, tech stack)
HubSpot MCP writes the enriched fields back to the contact record
Slack MCP posts a one-line summary to the assigned AE’s DM

The prompt looks like this:

Pull every contact created in HubSpot in the last hour. For each:
1. Enrich via Apollo (email, phone, LinkedIn)
2. Enrich via ZoomInfo (revenue, headcount, tech stack)
3. Write back to HubSpot contact properties
4. DM the assigned AE in Slack with a 2-line summary

If any step fails for a contact, log the failure and continue with the rest.

Time: 4 minutes for 20 new contacts. Manual time: 90 minutes.

Parallel chaining

The AI calls multiple servers at the same time, then joins the results.

Example: build a unified pipeline view.

HubSpot MCP returns deal stage, amount, and original source
Salesforce MCP returns the same deals from the system of record (when both run in parallel for SFDC-primary teams)
Google Ads MCP returns spend by campaign for the matched time window
LinkedIn Ads MCP returns spend by campaign for the matched time window
The AI joins all four datasets by UTM and campaign name

Sample prompt:

Pull all closed-won deals from Q1 in HubSpot. In parallel, pull Q1 spend from Google Ads and LinkedIn Ads. Join by UTM source and campaign. Output a table with columns: Channel, Campaign, Spend, Deals Won, Pipeline $, ROAS. Sort by ROAS descending.

Time: 30 seconds. Manual time: 3 to 4 hours of CSV exports and VLOOKUPs.

Conditional chaining

The AI decides which server to call based on what it finds. If A, do X. If B, do Y.

Example: tiered ABM response.

GA4 MCP checks engagement on the pricing page in the last 7 days
If engagement is high (3+ sessions), Apollo MCP looks up the buying committee
If engagement is medium, HubSpot MCP queues a nurture email instead
If engagement is low, do nothing
In all cases, Slack MCP posts the routing decision to #abm

For each Tier 1 ABM account, check pricing page sessions in GA4 over the past 7 days.
- If sessions >= 3: pull the buying committee from Apollo and create an SDR task in HubSpot
- If sessions = 1 or 2: enroll the contact in the "Pricing Page Visited" nurture flow in HubSpot
- If sessions = 0: skip
Post the count of accounts in each bucket to #abm in Slack.

Time: 10 minutes for 50 accounts. Manual time: half a day.

How chaining works under the hood

Three things make chaining reliable:

The AI keeps a “tool memory” within the chat. Numbers from step 1 stay available for step 4.
Each MCP server returns structured data (JSON), so the AI joins fields by key without parsing CSVs.
The AI handles partial failures. If 1 of 20 contacts fails enrichment, the chain continues for the other 19.

The big constraint: rate limits. If you chain 5 servers across 100 records, you might hit per-minute API caps on Apollo or HubSpot. The fix is batching, which the AI handles when you tell it to.

Sample batching instruction inside a prompt:

Process contacts in batches of 25. Wait 30 seconds between batches to respect Apollo's rate limit.

When to chain versus when to use a single server

Chain when:

The answer needs data from 2 or more tools (attribution, enrichment, routing)
The workflow has a clear hand-off between steps
The output is one consolidated answer, not separate reports

Use a single server when:

The question is fully answered inside one tool (“What is the open rate of our last LinkedIn campaign?”)
You’re exploring data and not sure what you need yet
Speed matters more than completeness

Action item: Map out one workflow you run weekly. Write the steps as a numbered list. Mark which tool answers each step. Where you have 2 or more tools, you have a chaining candidate.

How do B2B marketers run SEO and AEO workflows on MCP?

B2B SEO has a distinct shape: bottom-of-funnel comparison queries, alternatives content, and category-defining keywords. MCP makes the data side of B2B SEO conversational.

The five SEO MCP servers worth knowing:

Ahrefs MCP for backlinks, keyword data, traffic estimates, and Brand Radar AI mentions
Semrush MCP for competitor research and keyword difficulty (read-only, uses API credits)
Google Search Console MCP, free and open source, with 20 built-in tools
DataForSEO MCP for live SERP data, AI Overview detection, and SERP feature tracking
Frase MCP for read-write content optimization and AI visibility tracking across eight platforms

Six B2B SEO prompts to run this week:

“Find comparison keywords like ‘X vs Y’ where competitors rank top 3 and we don’t appear”
“Show alternatives queries (e.g. ‘alternatives to Salesforce’) with traffic but no ranking”
“Identify striking-distance keywords ranking 4 to 15 with buyer intent in the query”
“Pull our top 10 competitors’ newest blog posts and the keywords they target”
“Diagnose why our pricing page lost organic traffic last month”
“Find decision-stage queries (demo, pricing, buy) where we have low rankings”

Answer Engine Optimization (AEO) matters more in B2B than B2C. B2B buyers ask ChatGPT and Perplexity questions like “best CRM for mid-market SaaS” or “Salesforce alternatives under $50 per seat.” If your brand isn’t cited, the deal never reaches your funnel.

Tools like Frase MCP and Ahrefs Brand Radar track AI visibility across platforms. Sample prompt: “Check our brand visibility for ‘best B2B attribution platform’ across ChatGPT, Perplexity, Claude, Gemini, and Google AI Overviews. List which competitors get cited and the messaging angles they win on.” The server queries each platform and returns a citation breakdown.

Action item: Install the free Google Search Console MCP server this week. Run the comparison and alternatives queries above. Pick three bottom-funnel keywords to attack before Friday.

How does one prompt run an entire B2B campaign end to end?

Yes, with the right MCP stack. An ABM campaign normally needing a target list, an enrichment pass, ad creative, sequence drafts, and SDR briefs runs through one chat with Claude when six MCP servers are wired in.

The setup: HubSpot, Apollo or ZoomInfo, LinkedIn Ads, Outreach (or Salesforge), GA4, and Slack MCP servers connected to one Claude project. A single prompt kicks off the chain.

The prompt looks like this:

SYSTEM: You are a senior demand gen operator running an ABM launch.

<context>
Goal: launch a 6-week ABM campaign targeting our top 50 enterprise accounts
Budget: ${{TOTAL_BUDGET}} split 70% LinkedIn ABM display, 30% paid search
Target accounts: top 50 enterprise from our HubSpot 'Tier 1 ABM' list
Buyer committee: VP Marketing, VP RevOps, CMO
Offer: technical evaluation guide plus custom demo
Brand voice: {{BRAND_VOICE_NOTES}}
</context>

MUST follow these rules:
1. Pull the target account list from HubSpot before drafting anything
2. Enrich every account with recent funding, headcount, and tech stack via Apollo
3. Get human approval before activating any paid spend or sequence
4. Post a daily Slack summary to #abm-launches at 9am

Task: Plan, draft, and stage the campaign across HubSpot, Apollo, LinkedIn Ads, and Outreach.

Output:
- Enriched account list with key contacts per buying committee role
- LinkedIn ABM display creative brief and three ad variants
- Three SDR sequence variants for VP Marketing personas
- Budget split by channel with rationale
- SDR brief document with talking points per account
- Launch checklist for human approval

Claude runs the chain in order. Step one: query HubSpot for the Tier 1 account list and existing engagement history. Step two: enrich every account through Apollo with funding, headcount, and tech signals. Step three: identify the buying committee contacts per account. Step four: draft LinkedIn ABM display creative for each persona.

Step five: draft a 5-step Outreach sequence per persona. Step six: build SDR briefs with one-page account summaries. Step seven: stage every asset in draft state, never live. Step eight: post the launch checklist to Slack and wait.

The human reviews the checklist, approves or edits, and sends one final prompt: “Activate the approved variants.” Claude flips each campaign live. The whole chain takes 30 minutes instead of 3 weeks.

The agentic part comes after launch. Claude monitors the campaigns daily through the same MCP stack. If LinkedIn Ads CPL crosses the threshold, Claude pauses the offender. If an account shows three sessions on the pricing page (signal pulled from GA4), Claude flags the account for SDR outreach in Slack. If Outreach reply rates fall below 2%, Claude drafts a variant for review.

Three guardrails make this safe:

Write actions stay in draft mode until a human approves
Spend caps live in the prompt, not the AI’s discretion
Daily Slack summaries keep the team in the loop on every change

Action item: List the six tools you would need connected to run a full ABM campaign through one chat. Identify which already have official MCP servers and which need a Zapier MCP bridge.

How do teams scale MCP across an agency or RevOps team?

Scaling MCP follows three patterns: shared servers, prompt libraries, and custom internal MCP servers. Each one removes a different bottleneck.

Shared servers. One Claude Team workspace connects every member to the same MCP stack. The HubSpot admin authenticates once, then anyone on the account uses the connection. No per-user setup.

Prompt libraries. The five prompts your team runs weekly become saved templates. Store them in Notion or a shared doc. New hires copy and run on day one instead of learning every dashboard.

Custom internal MCP servers. The biggest scaling lever for agencies. Wrap your proprietary tools (a custom client reporting database, an internal margin calculator, a brand voice library) in your own MCP server. Every client account benefits from one build.

Real example from agency operators: a B2B demand gen agency built one internal MCP server pulling client pipeline data from HubSpot and Salesforce into a private Postgres database. Account managers ask Claude, “Which clients are pacing under quota this quarter and which channels are dragging?” and get an answer in five seconds. The same query took 45 minutes of CSV exports before MCP.

Five practical scaling moves:

Document which MCP servers are connected in a shared Notion doc
Set quarterly access reviews so disconnected accounts get cleaned up
Train new hires on three core prompts before tool training
Build one custom MCP server wrapping your most-queried internal data
Track time saved per workflow and report monthly to leadership

McKinsey research found agentic systems accelerate the creation and execution of marketing campaigns by 10 to 15 times by speeding up brainstorming, vetting, testing, and optimization. The catch: fewer than 10% of CMOs have deployed end-to-end workflows generating measurable value. The gap is implementation, not capability.

Action item: Audit your agency or team this week. Count how many client logins each account manager juggles. Pick the top three logins and check whether each tool ships an MCP server.

How do you build your own MCP server?

You build a custom MCP server when no existing one fits or when you want AI access to an internal tool. The official SDKs handle the hard parts.

The Python SDK and TypeScript SDK both expose three core primitives: tools, resources, and prompts. Tools are functions the AI runs, like “search contacts” or “create invoice.” Resources are read-only data sources. Prompts are reusable templates.

Quickest path with Python:

Install Python 3.10 or higher
Install the MCP package with pip install mcp
Create a server.py file
Define tools using Python decorators with type hints
Run with python server.py

A minimal server file looks like this:

from mcp.server.fastmcp import FastMCP

mcp = FastMCP("my-marketing-server")

@mcp.tool()
def get_campaign_status(campaign_id: str) -> str:
    """Get the current status of a marketing campaign by ID."""
    # Replace with your actual API call
    return f"Campaign {campaign_id} is active"

if __name__ == "__main__":
    mcp.run()

The FastMCP class reads your function names, type hints, and docstrings, then auto-generates the tool schema the AI needs. No JSON Schema writing required.

For TypeScript, the equivalent uses the @modelcontextprotocol/sdk package with Zod for input validation. Same pattern, different syntax.

Action item: Read the official MCP quickstart at modelcontextprotocol.io. Build the weather example server in 30 minutes to learn the pattern before tackling your own use case.

Which prompt helps you plan a custom MCP server?

Use this prompt to plan the tool list before writing any code. Paste into Claude or ChatGPT.

SYSTEM: You are an MCP server architect for marketing teams.

<context>
Tool to wrap: {{INTERNAL_TOOL_NAME}}
What the tool does: {{TOOL_PURPOSE}}
Common marketing tasks our team runs against this tool: {{TOP_TASKS}}
Who will use the AI assistant: {{USER_ROLES}}
</context>

MUST follow these rules:
1. Recommend 5 to 8 tools, no more
2. Each tool does one thing well, no kitchen-sink tools
3. Name each tool in verb_noun format (search_contacts, create_segment)
4. List required and optional parameters for each tool
5. Flag which tools need write access versus read-only

Task: Design the tool list for a custom MCP server that wraps {{INTERNAL_TOOL_NAME}}.

Output format:
- Tool name
- One-sentence description the AI reads to decide when to call it
- Parameters with types
- Read or write
- Sample prompt a user would type to trigger the tool

Run the prompt twice with the same inputs. Compare the two outputs and pick the cleaner tool list.

How do you host an MCP server?

Two hosting models exist: local and remote. Local servers run on the user’s machine and use stdio for transport. Remote servers run on the web and use HTTPS with the Streamable HTTP transport.

Local makes sense for personal tools and developer workflows. Remote makes sense when many team members need access or when the server connects to shared business data.

The four most common remote hosts:

Cloudflare Workers, free for 100,000 requests per day, near-zero cold starts
Vercel, integrates with Next.js projects, ships an mcp-handler package with built-in OAuth
Railway, container-based, $5 monthly free credit
FastMCP Cloud, a managed Python-first option with a free personal tier

Cloudflare offers a guide to deploy your own remote MCP server using Streamable HTTP transport, the current MCP specification standard. The basic flow: write your server code, install the Cloudflare CLI (Wrangler), deploy with wrangler deploy, and Cloudflare returns an HTTPS URL.

For Vercel, the mcp-handler npm package wraps the protocol so you focus on tool logic. Add OAuth with the withMcpAuth wrapper.

Action item: If you build a custom server, deploy to Cloudflare Workers first. Free tier covers most marketing teams, and the deploy takes under five minutes.

How do you add authentication to a remote MCP server?

Authentication matters when the server reads private data. Two patterns work: API keys for server-to-server calls, and OAuth 2.1 for user-facing servers.

For internal team servers, an API key passed in the X-API-Key header works fine. Store the key in environment variables, never in code.

For public servers like the HubSpot or Klaviyo style, use OAuth. You connect your MCP server with any OAuth provider supporting the OAuth 2.0 specification, including GitHub, Google, Slack, Stytch, Auth0, and WorkOS.

Cloudflare Workers ship a GitHub OAuth template. Run npm create cloudflare@latest — my-mcp-server-github-auth —template=cloudflare/ai/demos/remote-mcp-github-oauth and the scaffolding handles the auth flow.

For Claude Connectors specifically, allowlist both claude.ai and claude.com OAuth callback URLs in your provider. Missing one of these is a common reason Connector submissions get rejected.

What are the security risks of MCP?

MCP introduces three new attack types marketers need to know about: prompt injection, tool poisoning, and lookalike tools. None should stop you from using MCP, but each shapes how you connect servers.

Prompt injection tricks AI agents into executing hidden commands embedded in user inputs or external data sources. Tool poisoning embeds malicious instructions in tool metadata invisible to users but visible to AI models.

A real-world example: a malicious MCP server lists a tool called “send_email” with a description secretly instructing the AI to also forward all emails to an attacker address. The user sees a normal tool name. The AI sees the hidden instruction.

Five rules to stay safe:

Connect only servers from known publishers (HubSpot, Klaviyo, Anthropic, Cloudflare)
Avoid random GitHub MCP servers without a clear maintainer
Read the tool descriptions before approving any new server
Keep human approval on for write actions like sending emails or pausing campaigns
Audit which servers are connected each quarter and remove unused ones

The official MCP specification states “for trust and safety and security, there should always be a human in the loop with the ability to deny tool invocations.” Keep this loop closed.

Action item: Open your Claude Connectors settings today. List every connected MCP server. Disconnect any you do not recognize or no longer use.

What is the future of MCP for B2B marketing?

The trend is clear and one-way. Every major B2B SaaS platform is shipping or planning an MCP server. By early 2026, over 10,000 MCP servers exist with dedicated servers for every major B2B tool.

Three changes are coming over the next 12 months. First, write access expands. Most MCP servers today are read-only. HubSpot, Salesforce Agentforce, Apollo, and Outreach are all rolling out write operations. The AI will move from reporting to executing.

Second, MCP Apps. MCP Apps (formerly mcp-ui) is an official extension to the Model Context Protocol, formalized under the SEP-1865 specification in early 2026. MCP Apps let servers ship rendered widgets back to the chat: a deal pipeline view, an account scoring grid, a sequence editor. You’ll start interacting with these widgets directly inside Claude.

Third, agent-to-agent flows. AI agents will chain MCP calls across multiple servers without human prompting. The marketer sets the quarterly target, the agent runs the workflow loop. McKinsey research projects agentic AI will power as much as two-thirds of current marketing activities.

The teams getting ahead are the ones wiring up 4 or 5 servers this quarter, learning the prompt patterns, and building shared prompt libraries before competitors notice.

Final Takeaways

MCP is the connection layer for B2B marketing AI. One protocol replaces hundreds of one-off integrations across CRM, ad platforms, sales engagement, and intent data.

Your B2B starter stack: HubSpot or Salesforce, LinkedIn Ads, Apollo or ZoomInfo, GA4, and Slack. Install through Claude Connectors with OAuth in under 2 minutes per tool.

Custom MCP servers take 30 minutes to build with the Python or TypeScript SDK. Cloudflare Workers free tier (100,000 requests per day) covers most use cases.

Security risks are real but manageable. Connect trusted publishers only, keep human approval on write actions, and audit connections quarterly.

RevOps teams who set up MCP this quarter get faster pipeline reporting, sharper account attribution, and tighter SDR loops than teams still exporting CSVs.