Skip to module content
🧠AI Academy
0/91
Sign in
Module 06 Β· ~7 min

The Marketing MCP Stack

The seven-server reference architecture per engagement.

Reading progress
0/8 Β· 0%

The big idea

πŸ’‘Key idea
An April 2026 review of the full MCP directory reached a clear conclusion: of 1,000+ listed servers, roughly 25 are production-quality for marketing work β€” and the minimum viable deployment at most engagements is **seven**. **Stack sprawl** is the most common MCP failure mode. Every extra server is context tax (more always-visible tool descriptions), permission risk (more potential write surfaces), and demo fragility (one more thing that can break in front of a client). The seven core slots: **Analytics** (GA4), **CRM** (HubSpot), **Paid-search** (Google Ads official read + custom guarded write), **Paid-social** (Meta Ads), **Search/research** (Perplexity or Exa), **Content/ops** (Notion or the client's authoring surface), **Delivery** (Slack). Those seven slots cover the vast majority of engagement needs. Extend beyond them only when the engagement specifically requires it.
Quick check
1 question Β· instant feedback
0/1
  1. The seven-server rule says:

Numbers that matter

7
Minimum viable MCP deployment per engagement β€” the seven-server rule.
digitalapplied 25-servers review, Apr 2026
~25
Production-quality marketing MCP servers worth considering.
1,000+
Total listed servers β€” most are noise for a marketing engagement.
Quick check
1 question Β· instant feedback
0/1
  1. The delivery slot (Slack) is:

Deep dive

4/4 open

Each slot has a default and battle-tested alternates. Swap based on what the client already uses β€” don't add slots.

**Analytics:** GA4 (default) / Plausible, Mixpanel, PostHog

**CRM:** HubSpot β€” generally available, read + write on contacts, companies, deals, tickets, and activities / Salesforce, Pipedrive, Attio. Note: HubSpot MCP is a *CRM connector*, not a prospecting tool. It doesn't do sourcing or sequences.

**Paid-search:** Google Ads official server (read) + a custom mutate layer for write operations / Pipeboard, Adspirer

**Paid-social:** Meta Ads / LinkedIn Ads (required for any B2B engagement)

**Search / research:** Perplexity or Exa / Brave, Tavily, Firecrawl

**Content / ops:** Notion / Webflow, Contentful, WordPress, Linear, Figma

**Delivery:** Slack / Gmail

The April 2026 agency-stack review measured production quality across 25 candidate marketing servers and found that seven slots covered the vast majority of engagement needs.

The cost of sprawl compounds fast: every extra server adds always-visible tool descriptions to context, creates another credential to manage, introduces another potential write surface, and adds another point of failure during a demo.

The discipline is: **document each server's OAuth setup once as a client-installable runbook step, then add servers beyond the core only when the engagement specifically requires them.**

Runbook-once-reuse-forever is what makes this scale.

When the engagement explicitly requires outbound or lead-gen, add from this list β€” don't install them by default.

**Amplemarket** (Mar 2026) β€” the most complete single-platform outbound server. Read + write breadth is real.

**Apollo** (Feb 2026) β€” already widely used; OAuth-authenticated, free for paid Apollo customers. Pragmatic default for outbound.

**Clay** β€” orchestration hub plus Claygent for autonomous research. Credit-metered (~20–100 credits per action, 500 free). Build budget guardrails before using on client work.

**Instantly** β€” cold-email execution, 31 tools. Watch shared rate limits across workspace keys; bulk operations can stall mid-task.

**Smartlead** β€” deliverability diagnostics. SSE-only transport (no HTTP streaming).

**Salesforge Forge** β€” whole-stack outbound in one server.

**ZoomInfo GTM.AI** β€” generally available as of Jun 2026. Evaluate per-call caps (25/call) before committing to scale jobs.

**Explorium** β€” server-side batching for high-volume enrichment; the right choice when ZoomInfo's caps would cause context bloat.

Slack is output-only in the reference stack β€” the agent can't read channels, only post to them. That constraint makes it safe and easy to approve.

But it's essential. The delivery slot is what closes the gap between *insight generated* and *team action taken*. Morning performance briefs, anomaly alerts, weekly summaries β€” all of these land where the team already lives, without requiring them to log into a dashboard.

Skipping the delivery slot is the fastest way for an otherwise well-designed MCP stack to produce zero observable business impact. The insight runs, the agent produces output, and it disappears into a file nobody opens.

Slot seven is where value becomes visible.

Quick check
1 question Β· instant feedback
0/1
  1. HubSpot MCP is:

How to run it

  1. Deploy the seven core slots
    Analytics, CRM, paid-search, paid-social, research, content/ops, delivery.
  2. OAuth runbook per server
    Documented once, reused every engagement.
  3. Test-connectivity command
    One slash command that hits every server; logs the audit trail.
  4. Extension only on demand
    Add servers only when the engagement specifically requires them.
Quick check
1 question Β· instant feedback
0/1
  1. Adding servers beyond the seven core requires:

In the field

πŸ”¬Worked example
Lead-gen extension pack (only when the engagement requires): Amplemarket (most complete single-platform outbound, Mar 2026), Apollo (Feb 2026, OAuth, free for paid), Clay (orchestration + Claygent research, credit-metered), Instantly (cold-email exec, 31 tools), Smartlead (deliverability), Salesforge Forge (whole-stack), ZoomInfo GTM.AI (GA Jun 2026), Explorium (server-side batching for scale).
🚫When not to reach for it
Add servers beyond the core only when the engagement specifically requires them. Every additional server is context tax, permission risk, and one more thing that can break during a demo.
Quick check
1 question Β· instant feedback
0/1
  1. For B2B engagements the paid-social slot typically includes:

Pitfalls & takeaways

Failure modes

  • Stack sprawl β€” 15 servers installed, context bloated, half unused.
  • Skipping the delivery layer (Slack) β€” insights never reach the team.
  • Using HubSpot MCP as a prospecting tool β€” it's a CRM connector, not a sourcer.
  • One-runbook-fits-all β€” OAuth setup steps must be documented per server, per client.

Durable takeaways

  • Seven servers is the core; extend only on demand.
  • Runbook-once-reuse-forever is the discipline that scales.
  • The delivery layer (Slack) is what turns insight into action.
  • HubSpot MCP is a CRM connector β€” not a prospecting tool.

Do the work

πŸ‹οΈProve you learned it

Deploy the seven-server core for one client. For each server, document the OAuth setup as an installable runbook step. Verify every server responds via a single test-connectivity slash command, then log the audit trail.

0 chars
πŸ“¦Artifact to produce
Client MCP install runbook (7-server core + OAuth per-server steps) as a versioned skill.

Sources

  • Β· GMS Field Manual Β§6 (Marketing MCP Stack)