Intern Agent for Startup Ideas

An AI-powered agent that continuously generates, researches, and evaluates venture-backable startup ideas, saving the most promising ones as text files.

Features

• Autonomous Research: Uses Claude API with web search to research trends, markets, and competition in real-time
• Rigorous Evaluation: Only saves ideas that meet venture capital criteria (TAM $1B+, clear moat, timing, etc.)
• Continuous Operation: Runs 24/7 in Docker, generating ideas around the clock
• Enhanced Memory System: Tracks both approved and rejected ideas with reasons, compressing learnings to improve future evaluations
• Smart Deduplication: Avoids repetition by learning from both successes and failures
• Organized Storage: Saves each approved idea as a timestamped text file in the ideas/ directory

How It Works

1. Generate: Claude researches current trends and generates a startup idea
2. Research: Performs web searches to validate market size, competition, and timing
3. Evaluate: Applies VC criteria to determine if the idea is truly venture-backable
4. Learn: Tracks rejected ideas with reasons and compresses learnings when history grows large
5. Save: If approved, saves a detailed text file to ideas/ with justification, TAM, ICP, and competitive analysis
6. Loop: Continues immediately to the next iteration, learning from past successes and failures

Prerequisites

• Docker and Docker Compose (optional, for containerized deployment)
• Anthropic API key (Claude)
• Python 3.11+ (if running locally)

Setup

1. Clone/Download this repository

2. Set up environment variables

# Copy the example file
cp .env.example .env

# Edit .env with your credentials
nano .env

Fill in your API key:

ANTHROPIC_API_KEY=sk-ant-...

3. Create data directory (for Docker deployment)

mkdir -p data

The ideas will be saved to the ideas/ directory (automatically created on first run).

Running the Agent

Using Docker Compose (Recommended)

# Build and start the agent
docker-compose up -d

# View logs
docker-compose logs -f

# Stop the agent
docker-compose down

Using Docker directly

# Build the image
docker build -t startup-agent .

# Run the container
docker run -d \
  --name startup-agent \
  --env-file .env \
  -v $(pwd)/data:/app/data \
  startup-agent

# View logs
docker logs -f startup-agent

# Stop the container
docker stop startup-agent

Running locally (for testing)

# Install dependencies
pip install -r requirements.txt

# Run the agent
python agent.py

Viewing Generated Ideas

All approved ideas are saved as text files in the ideas/ directory:

# List all generated ideas
ls -lh ideas/

# View a specific idea
cat ideas/20251027_123456_AI_Governance_Platform.txt

# Monitor new ideas in real-time
watch -n 5 'ls -lh ideas/ | tail -10'

Configuration

Adjusting Research Depth

Edit agent.py line 28 to change the number of web searches per idea:

MAX_WEB_SEARCHES = 10  # Increase for deeper research, decrease for faster generation

Adjusting Learning Compression

Edit agent.py line 29 to change when rejected ideas get compressed into learnings:

COMPRESSION_THRESHOLD = 100  # Number of rejected ideas before compression runs

Modifying Evaluation Criteria

Edit the system prompt in agent.py to adjust what makes an idea "venture-backable."

Monitoring

View real-time logs

docker-compose logs -f

Check idea history

cat data/ideas_history.json

Restart the agent

docker-compose restart

Troubleshooting

"ANTHROPIC_API_KEY not set" error

• Ensure your .env file exists in the project root
• Check that ANTHROPIC_API_KEY is set correctly
• Verify the API key is active at https://console.anthropic.com/

"Anthropic API error"

• Verify your ANTHROPIC_API_KEY is correct
• Check your API usage limits at https://console.anthropic.com/
• Web search costs $10 per 1,000 searches plus token costs

No ideas being saved

• Check logs: docker-compose logs -f or cat agent.log
• The agent is highly selective - it may take several cycles before approving an idea
• Check the ideas/ directory for newly saved files
• Review rejected ideas in ideas_history.json to see what's being filtered out

Container keeps restarting

• Check logs: docker logs startup-idea-agent
• Verify all environment variables are set
• Ensure the data directory exists and is writable
• Ensure the ideas directory can be created

Cost Estimation

Claude API Costs

• Web search: $10 per 1,000 searches
• Tokens: ~$3 per million input tokens, ~$15 per million output tokens
• Estimated: $0.50-2.00 per idea (depending on research depth)

With 30-60 ideas per day, expect roughly $15-60/day in API costs.

Architecture

agent.py              # Main agent loop and idea generation
├─ generate_and_evaluate_idea()  # Core logic with web search
├─ load_history()     # Load approved/rejected ideas and learnings
├─ save_history()     # Persist history with timestamps
├─ save_idea_to_file()  # Save approved ideas as text files
└─ compress_learnings()  # Compress rejected ideas into insights

ideas_history.json    # Enhanced history tracking (auto-generated)
├─ approved_ideas[]   # Titles and keywords of approved ideas
├─ rejected_ideas[]   # Rejected ideas with reasons and timestamps
├─ compressed_learnings  # AI-generated insights from rejections
└─ last_updated       # Timestamp of last update

ideas/                # Directory with approved idea files (auto-created)
└─ YYYYMMDD_HHMMSS_Idea_Title.txt  # Timestamped idea files

Memory System

The agent now features an enhanced memory system that learns from both successes and failures:

• Approved Ideas: Tracked by title and keywords to avoid repetition
• Rejected Ideas: Stored with detailed reasons (market issues, competition, TAM concerns, etc.)
• Compressed Learnings: When rejected ideas exceed the threshold (default: 100), Claude automatically summarizes common rejection patterns into actionable insights
• Context Integration: Both recent history and compressed learnings are included in prompts to improve future evaluations

Customization Ideas

• Add Slack/Discord notifications when new ideas are saved
• Parse text files into a database for analysis and visualization
• Build a web dashboard to browse and filter saved ideas
• Integrate with Airtable or Notion for idea management
• Add semantic similarity checking using embeddings
• Filter by specific industries or technologies
• Add email notifications (using email_sender.py as a starting point)
• Export ideas to PDF with better formatting

License

MIT

Support

For issues or questions, contact team@heysanctum.com