Video summary

Claude Code & MCPs built my $145K marketing machine

Main summary

Key takeaways

Business

Business-specific summary: “Claude Code & MCPs built my $145K marketing machine”

Core idea (what they built)

They describe a GTM engineering workflow where AI agents—using Claude Code and MCPs—do the “middle work” in marketing and sales without manual keyboard work, running 24/7.

This system performs tasks such as:

  • Research
  • Lead sourcing
  • Creative generation
  • Publishing
  • Measurement
  • Optimization

They demonstrate a recurring loop:

  1. Generate large volumes of ad creative from pain-point research
  2. Upload/batch the creative to Facebook
  3. Pull live performance data via a data pipeline / warehouse
  4. Turn off poor performers and promote winners by reallocating budgets

Tools + systems mentioned (stack-level view)

Agent + integration layer

  • Claude Code / “vibe coding” + skills/agents
  • MCP (Model Context Protocol) for connecting to data/outputs (e.g., “Graft MCP”)

Outbound / lead generation + messaging

  • PhantomBuster (scrape LinkedIn engagers; use PhantomBuster API)
  • Instantly (cold email campaigns; used in multi-step outreach)
  • MillionVerifier (email verification)
  • Apollo API (enrich LinkedIn profiles)

Creative + assets

  • Facebook Ads API (bulk create + draft upload; also pause/turn off ads)
  • Perplexity API (research pain points from sources like Reddit)
  • Verifonic (listed in the stack)
  • Railway (deploy servers so agents can run persistently)
  • React + HTML Canvas (build UIs and export creatives as PNGs)
  • HeyGen API (port static creative into UGC/video formats)

Analytics + dashboards

  • GA4 via MCP (example: morning KPI checks)
  • “Graft” + Looker Studio / BI-style dashboards (scorecards, line charts, demographic views)

Playbooks / frameworks / process patterns

  • Agentized GTM pipeline (end-to-end automation)

    • Input: voice + API keys + templates/skills in a repo
    • Automate: research → generate assets → publish → analyze → optimize
    • Execution: multiple agents “jockeying” across windows/tasks

  • Creative testing loop (ad optimization system)

    • Batch create mass variants from pain-point-driven angles
    • Upload as drafts into a Facebook ad set
    • Measure performance signals (clicks, spend, CPM, CPC, etc.)
    • Kill switch: pause/turn off low performers
    • Scale winners: move winning creatives into new ad sets with dedicated budgets
    • Run on a schedule (e.g., daily cron job)

  • Data-pipeline requirement

    • Emphasis that you need a data warehouse / data pipeline to analyze at scale.
    • MCP reads live data endpoints feeding the warehouse, rather than relying only on UI scraping.

  • “Personal software” / repo-first ops

    • Build skills inside a persistent code folder (“repo”)
    • Centralize API keys in an environment file
    • Optionally add transcription tools to accelerate workflow creation

Concrete examples / case demos (what they actually did)

1) Facebook ad “autopilot” demo (mass creative + optimization)

  • Spin up 10 Claude Code instances
  • Workflow includes:
    • Create 100 Facebook ads (image size referenced as “1080x1080”)
    • Publish to Facebook
    • Build a performance dashboard
    • Use Claude to analyze results
    • Turn off low performers and promote winners into a new ad set with its own budget
  • Claimed speed: “in ~30 minutes” for the described loop (at demo scale)

2) Ad creative generator from public pain points

  • Use Perplexity API to pull research sources (e.g., Reddit, other social platforms)
  • Extract:
    • pain points
    • desired outcomes growth marketers want (from BI/reporting tools)
  • Generate ad copy variations and creative templates at scale
  • They describe two approaches:
    • Generate maximum variations quickly (cheap tokens) to find “winners”
    • Spend more compute/tokens on higher-quality “scroll-stopping” variants once angles/format winners are identified
    • (They mention alternatives like Nano Banana Pro, but focus on code/template generation for flexibility and brand control.)

3) Podcast outreach “scrape → verify → email” pipeline

  • Scrape podcast host emails using a verification flow:
    • Refonic (via “Rephonic API,” as described)
    • Verify emails with MillionVerifier
    • Send outreach via Instantly
  • Earlier example: scrape marketing category podcasts → cold email to get booked (reported performance “way better than expected”)

4) LinkedIn engagement → enrichment → cold email

  • Workflow:
    • Slack command: user posts a LinkedIn URL
    • PhantomBuster extracts people who engaged with the LinkedIn post
    • Apollo API enriches LinkedIn profiles
    • Verify emails with MillionVerifier
    • Add leads to an Instantly campaign

5) Tracking dashboard creation

They build dashboards from Facebook ad data including:

  • Line chart for clicks over time
  • Cost and CPC lines
  • Scorecards for total spend and total clicks
  • Demographic breakdown by age categories using impression data

6) Optimization using CPM via MCP + “turn off losers”

  • Use “Graft MCP” to pull CPM data from Facebook ads
  • Pause ads with the highest CPM (treated as “low performers” in their demo logic)

7) On-demand “ephemeral” data workflows using Railway + Postgres

They replaced hours of manual Excel/pivot work with:

  • Railway-spun up Postgres pushed from a URL
  • Analysis + processing with Claude
  • Shut down the database to save cost/time

Claimed outcome:

  • ~5 hours → 20–30 minutes

Metrics & KPIs mentioned (and how they’re used)

Marketing performance KPIs

  • CPM (used to identify low performers and pause them)
  • CPC
  • Clicks over time
  • Total spend
  • Total clicks
  • Impressions by age category
  • “CPM data” + dashboards drive decisions to disable ads

Operational / product KPIs (implied)

  • Website traffic / “new users to homepage” (example morning check via GA4 MCP)
  • Business ROI framed in ad-economics terms: “put in a dollar, get $3 out”
  • Time-to-cycle improvements:
    • Create 100 ads + analyze + turn off low performers + promote winners in ~30 minutes
    • Data cleaning 5 hours → 20–30 minutes

No explicit CAC/LTV/churn targets were stated numerically; emphasis is on execution KPIs like spend, CPC/CPM, clicks, and traffic.

Actionable recommendations they imply (how to adopt)

  • Start with a repo/folder that stores:
    • skills/agents
    • an environment file with API keys for your marketing stack
  • Build workflows around specific jobs-to-be-done, not “one agent with everything”
  • Use batch creation:
    • generate many variations quickly
    • use performance data to identify winners
    • reallocate budget to better-performing ad sets
  • Deploy agents to run “in perpetuity” (e.g., to Railway with an API-backed server)
  • Build dashboards that provide decision-grade visibility:
    • clicks, CPC, CPM, spend, demographic views
  • Use MCP/data feeds that connect to live endpoints + data warehouse, rather than only UI interactions

Investing/markets note (high-level only)

The talk is heavily execution-focused. They briefly frame a future where marketing becomes increasingly autonomous (“autonomous marketing”), which could disrupt headcount but also create new roles in agent/ops design.

Presenters / sources

  • Presenter / guest: Cody Schneider (friend; described as a “legend” in vibe marketing/growth marketing)
  • Host: Unnamed speaker/interviewer (not clearly identified in subtitles)

Original video