Promo Kit

Finished reference project for the MCP build night. This Manufact/mcp-use app creates a complete promo kit from a short campaign brief.

It combines:

• Exa for web research and source summaries
• Unsplash for campaign visuals by default
• optional fal.ai support for generated poster images
• ElevenLabs for voice ad generation
• Langfuse for trace and score observability
• an LLM-as-a-judge style evaluator for output quality
• MCP as the agent tool interface

Setup

Prerequisite: Node.js 22 or newer.

npm install
cp .env.example .env

Add your workshop credit keys to .env:

PORT=3000
MCP_URL=http://localhost:3000
IMAGE_PROVIDER=unsplash
EXA_API_KEY=...
UNSPLASH_ACCESS_KEY=...
FAL_KEY=...
ELEVENLABS_API_KEY=...
ELEVENLABS_VOICE_ID=TX3LPaxmHKxFdv7VOQHJ
JUDGE_PROVIDER=heuristic
OPENAI_API_KEY=
JUDGE_MODEL=gpt-4o-mini
LANGFUSE_PUBLIC_KEY=
LANGFUSE_SECRET_KEY=
LANGFUSE_BASE_URL=https://cloud.langfuse.com

ELEVENLABS_VOICE_ID is optional. The example value is a premade voice that worked with the free-plan key during testing. FAL_KEY is optional unless you set IMAGE_PROVIDER=fal. Langfuse and OpenAI are optional. Without them, the benchmark still works with a local heuristic judge and returns langfuse.sent: false. When Langfuse keys are missing, langfuse.dryRun: true shows the trace name and score names that would be sent. The app loads .env automatically on Node 22+, and direnv-exported variables still work.

If you use direnv, put those exports in .envrc.local instead and run:

direnv allow

Start the local server:

npm run dev

In another terminal, run the preflight smoke test:

npm run smoke

To verify the Langfuse request shape without real Langfuse keys:

npm run test:langfuse

Open the inspector:

http://localhost:3000/inspector

Local MCP endpoint:

http://localhost:3000/mcp

If mcp-use reports a different port because 3000 is busy, set PORT in .envrc.local and update mcp.json / .mcp.json to match for that machine.

Optional Web UI

The MCP inspector is a developer console, not the product UI. This repo also includes a tiny local web app that calls the same MCP tools through a local proxy, so you can showcase the MCP server and a real client side by side.

Start the MCP server first:

PORT=3022 MCP_URL=http://localhost:3022 npm run dev

Then start the web UI in another terminal:

MCP_SERVER_URL=http://localhost:3022 npm run web

Open:

http://localhost:5174

The UI can:

• run check_setup
• load list_demo_presets
• call run_demo_preset
• call create_and_evaluate_promo_kit
• call run_benchmark_suite
• call research_market
• call generate_voiceover

This is intentionally small enough for attendees to modify with Cursor during the workshop.

To point the local UI at a deployed MCP server instead of your local server:

MCP_SERVER_URL=https://fast-zero-4d5xn.run.mcp-use.com npm run web

Cursor Demo

Use mcp.json to connect Cursor to the local server:

{
  "mcpServers": {
    "promo-kit-mcp": {
      "url": "http://localhost:3000/mcp"
    }
  }
}

Then ask Cursor Agent:

Use get_workshop_flow first, then check_setup.
After that, use run_demo_preset with cursor-build-night-padova.
Show the promo kit, the judge score, and whether Langfuse received the trace.

Or, for the shortest possible demo:

Use run_demo_preset with cursor-build-night-padova.
Show the promo kit, the judge score, and whether Langfuse received the trace.

Tools

• check_setup()
• get_workshop_flow()
• research_market(topic, audience, location, maxResults)
• generate_poster(brief, visualStyle, format)
• generate_voiceover(script, voiceId, language)
• create_promo_kit(topic, audience, location, tone)
• evaluate_promo_kit(topic, audience, location, promoKitJson)
• create_and_evaluate_promo_kit(topic, audience, location, tone)
• run_benchmark_suite(topic, audience, location, tones)
• list_demo_presets()
• run_demo_preset(preset)

Expected Output

create_promo_kit returns:

• title and positioning
• three social captions
• Exa-backed research with source links
• Unsplash image URL with attribution, or fal.ai poster prompt and image URL when IMAGE_PROVIDER=fal
• ElevenLabs voiceover script, plus an audio data URL when the current key and voice have enough API access

create_and_evaluate_promo_kit returns:

• promoKit: the generated campaign kit
• evaluation: judge, overall score, rubric scores, strengths, and improvements
• langfuse: whether the trace and scores were sent to Langfuse, plus dry-run trace and score metadata when keys are missing

run_benchmark_suite returns:

• comparison: ranked rows for two or three tone variants
• winner: the winning variant ID
• candidates: each generated promo kit, judge evaluation, and Langfuse trace metadata

Demo Prompts

Run get_workshop_flow and explain the live demo sequence.

Run check_setup and tell me what is ready for the workshop.

Create a promo kit for a student AI build night in Rome.

Create and evaluate a promo kit for a Cursor build night in Padova for developers.

Run the cursor-build-night-padova preset and explain the judge scores.

Run a benchmark suite for a Cursor build night in Padova for developers.
Compare energetic and practical, friendly and beginner-safe, and bold and urgent tones.
Show the winner, the comparison table, and the Langfuse trace IDs.

Preset IDs:

• cursor-build-night-padova
• student-ai-build-night-rome
• matcha-cafe-university
• indie-game-tournament

Create a promo kit for a matcha cafe opening near a university.

Create a promo kit for an indie game tournament this weekend.

Troubleshooting

Missing API key:

EXA_API_KEY is required for Exa. Copy .env.example to .env and add your workshop credit key.

Fix: add the missing key to .env and restart npm run dev.

fal.ai returns no image URL:

fal.ai returned no image URL.

Fix: retry with a shorter prompt or check fal.ai credits.

Unsplash returns no image:

Unsplash returned no image

Fix: use a simpler campaign brief or visual style.

ElevenLabs returns an auth or quota error:

ElevenLabs TTS failed (401 or 429)

The server now keeps the promo kit usable when TTS is unavailable. The voiceover object returns status: "unavailable", keeps the script, and includes the provider error for troubleshooting.

Fix: verify the API key, voice ID, account tier, and remaining credits.

Langfuse says sent: false:

langfuse: { "enabled": false, "sent": false, "dryRun": true }

Fix: set LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, and LANGFUSE_BASE_URL, then restart npm run dev.

For the workshop, dryRun: true is still useful: show traceName, scoreNames, and scoreCount to explain exactly what the app would send to Langfuse once keys are configured.

OpenAI judge is not configured:

judge: "heuristic"

This is expected for the workshop. Set JUDGE_PROVIDER=openai and OPENAI_API_KEY only if you want a live LLM judge instead of the deterministic rubric.

Check setup before a live demo:

check_setup()

This reports which providers are configured without exposing key values. Exa and the selected image provider are required for the finished live demo. ElevenLabs, OpenAI judge, and Langfuse can be unavailable without breaking the main promo kit flow.

Langfuse implementation note:

This repo keeps the workshop dependency surface small by sending traces and numeric scores through Langfuse's public HTTP API from src/providers/langfuse.ts. The same flow can be swapped to the Langfuse SDK later without changing the MCP tool contract.

Build warning about large chunks:

Some chunks are larger than 1024 kB

This comes from the mcp-apps widget bundle and does not block the local workshop demo.

Optional Distribution Story

This repo includes local wrapper files for:

• Cursor: .cursor-plugin/plugin.json and mcp.json
• Claude Code: .claude-plugin/plugin.json and .mcp.json
• Codex: .codex-plugin/plugin.json and .mcp.json

The important idea for attendees: the MCP server is the product, and plugin or connector marketplaces are the distribution layer.

Workshop Narrative

1. The agent calls MCP tools instead of only writing text.
2. Exa grounds the campaign in current web context.
3. Unsplash supplies a visual asset with attribution.
4. ElevenLabs attempts a voiceover; if account limits block audio, the workflow still keeps the script.
5. The judge scores the output with a rubric.
6. Langfuse stores the trace and numeric scores when keys are configured.

Manufact Cloud

Deployment is optional for the local workshop:

npm run deploy

Do not commit real API keys, Manufact device codes, generated images, or generated audio files.