The Content Plugin

🎥 Video lesson — video not yet migrated (source Skool video id: 2529589855534d1ea170aa30466203d5).

MCP setup, 6 agents, and the full content pipeline — from competitive research to a scheduled Instagram post.

CW-6 gets the content plugin running end to end. You'll configure three MCP servers in Claude Desktop — YouTube for competitive research, Buffer for publishing, fal-ai for image generation — learn how to handle API keys safely, and run the full setup wizard to configure the plugin to your brand. Then you'll see the 6-agent pipeline in action: research, concept development, carousel generation, and scheduling a live post to Buffer — all without leaving Claude.

The lesson closes with a practical note on Cowork's sandbox limits and how to work around them.

What you'll set up → YouTube MCP — competitive research and video stats via Google Cloud Console API → Buffer MCP — scheduling and publishing across 13+ social platforms → fal-ai MCP — image generation (Nano Banana 2 for carousels and thumbnails) → API key security — what they are, how to scope them, what to do if one leaks → Content plugin setup wizard — brand voice, ICP, platform config in 8 sections

What you'll learn → What the content plugin does — 6 agents, one shared project folder → How the setup wizard builds your context folder (standalone / projects / context) → How agents hand off work through the project folder, not conversation memory → Which MCP powers which agent → How to work around Cowork sandbox limits for image downloads → Cowork vs Claude Code — sandboxed safety vs open-ended control

Next up: CW-7 — What's Next

PART 2 — LESSON BODY (full markdown, paste into Skool lesson page)

The Content Plugin

This lesson has two halves. First: set up the MCP servers the content plugin needs and configure it to your brand. Second: run the pipeline — competitive research, content ideation, Instagram carousel, live Buffer scheduling.

Part 1: What is an MCP?

Claude in Cowork is sandboxed. It can't make raw HTTP calls to YouTube, Buffer, or any other external service. MCP servers (Model Context Protocol) are the bridge — controlled connections that give Claude approved access to specific services.

Without an MCP: Claude knows about YouTube but can't touch it. With the YouTube MCP: Claude can search videos, pull channel stats, and fetch metadata.

Each MCP runs locally on your machine via Claude Desktop. You control what's connected. Each has its own API key, scoped to one service — a leak in one doesn't touch the others.

Setting up the MCPs

Open Claude Desktop. Go to Settings → Developer → Edit Config. This opens claude_desktop_config.json in your text editor. Add each of the three servers below inside the mcpServers object, then save and restart Claude Desktop.

YouTube MCP

"youtube": {
  "command": "npx",
  "args": ["-y", "youtube-data-mcp-server"],
  "env": {
    "YOUTUBE_API_KEY": "YOUR_API_KEY_HERE",
    "YOUTUBE_TRANSCRIPT_LANG": "en"
  }
}

Used by Content Strategist for competitive research, video stats, and keyword trends.

Getting your YouTube API key:

  1. Go to console.cloud.google.com
  2. Create a new project (name it something relevant to your use case)
  3. Go to APIs & Services and search for YouTube Data API v3 → click Enable
  4. Go to Credentials → Create Credentials → API Key
  5. Give the key a name, then restrict it to YouTube Data API v3 only — this is important. Without restricting it, the key could potentially access Google Drive, Sheets, and other Google services
  6. Copy the key and paste it into the config above

Buffer MCP

"buffer": {
  "command": "npx",
  "args": ["-y", "@damusix/buffer-mcp"],
  "env": {
    "BUFFER_ACCESS_TOKEN": "YOUR_ACCESS_TOKEN_HERE"
  }
}

Used by Publisher for scheduling and publishing across 13+ social platforms.

Getting your Buffer access token:

  1. Log into Buffer → go to your organisation → Settings → API
  2. Create a new key — name it something like "Claude Desktop [your name]"
  3. Set the expiry. Note: if you choose 30 days, you'll need to repeat this process after it expires
  4. Copy the token and paste it into the config above

fal-ai MCP

"fal-ai": {
  "command": "uvx",
  "args": ["--from", "fal-mcp-server", "fal-mcp"],
  "env": {
    "FAL_KEY": "YOUR_FAL_KEY_HERE"
  }
}

Used by Creative Director for image generation (Nano Banana 2 is the recommended model for carousels and thumbnails).

Getting your fal-ai API key:

  1. Go to fal.ai and sign up
  2. Go to API Keys → Add Key
  3. Give it a description (e.g. "Content plugin demo")
  4. Copy the key and paste it into the config above

Important — disable auto top-up: fal-ai is usage-based. Go into your fal.ai account and turn off auto top-up. This protects you in the unlikely event your API key gets leaked or Claude makes unexpected extra API calls — your account can only spend down to zero, not rack up an unbounded bill. Top up manually once you're comfortable with how much you're using.

API keys: handle with care

An API key is a password for a service. Whoever has your key can use the service in your name — and you pay the bill.

  • Never paste a key into a chat, document, or screenshot
  • Never commit a key to a GitHub repository — even a private one
  • Try not to show them on screen during recordings (delete before you finish)
  • If a key is ever exposed, rotate it immediately in the provider's dashboard
  • Each MCP's key is scoped to one service — a leaked YouTube key can't touch your Buffer account

After saving the config, restart Claude Desktop. Go back to Developer and confirm each server shows a running status. If there's an error, it'll show there.

Verifying the MCPs are working

Once running, you can do a quick test by asking Claude:

"With the YouTube Data MCP, can you make a quick connection to test that it's working?"

Expand the tool call in the response — you'll see the request and response. Do the same for Buffer and fal-ai. If any of them error, check the config JSON for a missing comma or mismatched key name.

You can also view your MCPs in Cowork by going to Connectors in the sidebar. Each connected MCP will show its available tools. You can toggle individual tool permissions here — if a tool can delete data and you never want Claude to use it, you can block it entirely. For this course, leave everything on needs approval — Claude will ask before executing anything.

Part 2: The Content Plugin

The content plugin is a 6-agent production pipeline. It takes you from competitive research to scheduled posts across 13+ platforms. Each agent owns one stage of the workflow.

The 6 agents

Research & Creation:

  • Content Strategist — competitive research, trends, content ideation. Uses YouTube MCP
  • Copywriter — video scripts, short form scripts, LinkedIn, X, blogs, email campaigns, diagrams
  • Creative Director — thumbnails, carousels, visual assets via fal-ai (Nano Banana 2)

Quality & Distribution:

  • Editor — optional quality review: brand voice, ICP relevance, value delivery
  • Publisher — content calendar, scheduling via Buffer MCP, post-publish pipeline
  • Autopilot — daily LinkedIn + X drafts, 3x/week Instagram carousels, review queue. Build to this once the manual pipeline is working reliably

How the pipeline flows

Agents share state through a shared project folder — not conversation memory. Each agent reads the previous agent's output as files. This is what lets you start a new Cowork tab, run a different agent, and have it pick up exactly where the last one left off.

  1. Content Strategist researches → saves a content brief and competitive research report to the project folder
  2. Copywriter reads the brief → writes scripts and post copy
  3. Creative Director reads the script → generates carousels and thumbnails
  4. Editor reviews output (optional) → scores for brand fit, ICP relevance, value
  5. Publisher schedules approved content → posts via Buffer
  6. Autopilot runs independently — generates daily drafts and queues them for approval

Which MCP powers which agent

  • YouTube MCP → Content Strategist: competitive research, keyword trends, video metadata
  • fal-ai MCP → Creative Director: image generation for carousels and thumbnails
  • Buffer MCP → Publisher: schedules and publishes across LinkedIn, Instagram, X, TikTok, and more

Setting up the plugin for your brand

Before running any content agents, run the Setup skill. This creates your context folder — the persistent brain the plugin reads every time it runs.

  1. In Cowork, click the folder dropdown and choose your content folder (or create one — e.g. Documents/cowork-context/content-plugin)
  2. Run the Setup skill and type SW to launch the full setup wizard
  3. It'll walk you through 8 sections — brand voice, ICP, platform config, tone, visual identity, etc.
  4. At the end, it creates a structured folder with three subdirectories:
    • standalone/ — for one-off tasks that don't belong to a project (e.g. a quick carousel)
    • projects/ — for full pipeline runs (research → scripts → visuals → scheduled)
    • context/ — your brand config files. The agents read from here every run

Every time you open a new Cowork tab to use this plugin, point it at this same folder. Claude reads the context files fresh each session — it's not loaded into its memory, it reads the files.

Running the pipeline

Step 1 — Content Strategist: competitive research

Run the Content Strategist skill and create a new project (give it a name relevant to the content you're making). The agent will:

  • Ask about your topic or video idea
  • Search YouTube for competitor videos in your niche
  • Pull stats (views, engagement rates, upload cadence)
  • Build a competitive research report saved to projects/[name]/strategist/research/

Note: in the Cowork sandbox, YouTube transcripts require a separate API endpoint that isn't accessible by default. The MCP gets you video metadata and stats — enough for solid research. If you want full transcripts, you can add the endpoint URL to Settings → Capabilities → Trusted Domains in Cowork.

Step 2 — Content Strategist: content ideation

Run the Content Strategist again in a new tab. It'll detect your existing project and ask whether to resume it. Choose resume. Then run the Content Ideation workflow:

  • It reads your competitive research report automatically
  • Asks whether you've already filmed (if yes, you can drop a transcript and it'll align the concept to what you actually said)
  • Develops angles for each content format: YouTube, short form, LinkedIn, etc.
  • Saves a concept brief to projects/[name]/strategist/ideation/

Step 3 — Creative Director: Instagram carousel

Run the Creative Director skill. Choose Visual Asset Creation → Instagram Carousel. It'll:

  • Read your brand guidelines and Instagram colour config
  • Ask how many slides, the hook style, and the topic
  • Generate a slide plan with headlines, layout, and colour choices
  • Generate each slide image via fal-ai (Nano Banana 2)

Cowork image download workaround: In the Cowork sandbox, fal-ai can generate images but can't download them directly into your folder by default. To fix this:

  1. Go to Cowork Settings → Capabilities
  2. Set network access to Package managers only (recommended baseline)
  3. Add two trusted domains: rest.fal.ai and v3b.fal.media

With those domains allowed, the Creative Director can pull generated images directly into your project folder. Without them, you can still grab images manually from fal.ai → Recent History in your browser.

Step 4 — Publisher: schedule to Buffer

Run the Publisher skill and choose Schedule and Publish. Reference the carousel you just built. The agent will:

  • Find your Instagram carousel in the project folder
  • Read the concept brief to generate a caption and hashtags
  • Ask when to schedule it
  • Create the post via Buffer MCP

Check your Buffer calendar — you should see the post scheduled at the time you specified.

Cowork vs Claude Code

The content plugin works in both environments. Same plugin, different tradeoffs.

Claude Cowork:

  • Sandboxed — Claude can only use approved MCPs and trusted domains
  • Team-shareable, admin-controlled
  • No terminal required
  • Lower risk, faster setup
  • Works for: daily content production, team deployment, clients

Claude Code:

  • Open — can reach filesystem, run scripts, install packages, call any API
  • Per-person config, manual setup
  • Terminal required
  • More power, more responsibility
  • Works for: builders who want hooks, Python scripts, full filesystem access

Roughly 80–90% of the efficiency gains are achievable inside Cowork. Claude Code is worth learning if you want to go deeper — hooks, automation scripts, custom tool chains — but it's not where most people should start.

For this course, we use Cowork. It's controlled, shareable, and covers the full pipeline without a terminal.

What's next

The plugin is configured. You've run the full pipeline — research to scheduled post. From here, the path is:

  1. Run it a few more times manually to see where your outputs need tuning
  2. Edit the context files (brand voice, ICP, platform config) to sharpen the results
  3. Once it's running reliably, engage Autopilot to start scheduling daily content without manual input

→ Next: CW-7 — What's Next

← Previous
Next →