7 Golems + Infrastructure, 3 Environments

Golems is a Bun workspace monorepo with 11 packages — 7 golems (1 orchestrator + 6 domain experts) plus shared infrastructure. Work splits between your local Mac (cognitive tasks), Railway cloud (data collection), and Vercel (web dashboard).

Mac = Brain, Railway = Body

Cloud Worker Schedule

The Railway cloud worker runs these jobs on a timer:

Cloud jobs use Gemini 2.5 Flash-Lite (free tier) for cost efficiency. Each job publishes events to Supabase that trigger Mac-side Golems.

Local Services (Mac)

Your Mac runs these always-on services:

The local services have direct compute access — they run local GLM-4.7-Flash (via Ollama) or cloud models when needed.

Event Flow

When an email arrives:

Environment Variables (Dual Mode)

Golems supports dual mode — run cloud or local via three env vars:

# LLM Backend: where LLM calls happen
export LLM_BACKEND=gemini     # Cloud: Gemini Flash-Lite (free, default)
export LLM_BACKEND=ollama     # Local: Ollama on Mac (for testing)
export LLM_BACKEND=haiku      # Cloud: Haiku (paid fallback, optional)
 
# State Storage: where data lives
export STATE_BACKEND=supabase # Cloud: Supabase database
export STATE_BACKEND=file     # Local: ~/.brainlayer/
 
# Notifications: where Telegram messages go
export TELEGRAM_MODE=direct   # Cloud worker sends directly
export TELEGRAM_MODE=local    # Mac notifier (HTTP) sends

Switching Modes

Full Cloud Mode (Production)

export LLM_BACKEND=gemini
export STATE_BACKEND=supabase
export TELEGRAM_MODE=direct
# Deploy to Railway, monitor /api/usage for token counts

Full Local Mode (Testing)

export LLM_BACKEND=ollama    # Run: ollama pull mistral
export STATE_BACKEND=file
export TELEGRAM_MODE=local
# Run Mac services: bun src/telegram-bot.ts

Hybrid Mode (Development)

export LLM_BACKEND=gemini     # Use cloud LLM (free Gemini Flash-Lite)
export STATE_BACKEND=file     # Store locally for iteration
export TELEGRAM_MODE=local    # Debug Telegram messages
# Perfect for testing new features before cloud deploy

Rollback

If something breaks in cloud, roll back in under 1 minute:

# Switch back to local-only (everything still works)
export LLM_BACKEND=ollama
export STATE_BACKEND=file
export TELEGRAM_MODE=local
 
# Restart Mac services
golems latest
 
# Check status
golems status

No data loss, no disruption. The state in Supabase is still there for when you re-enable cloud.

API Cost Tracking

All LLM calls are logged to a JSONL file:

# Location (Mac):
cat ~/.brainlayer/api_costs.jsonl
 
# Location (Cloud):
curl https://your-service.up.railway.app/usage

Format:

{"timestamp": "2026-02-06T10:30:45Z", "model": "claude-haiku-4-5-20251001", "source": "email-poller", "input_tokens": 1240, "output_tokens": 340, "cost_usd": 0.002352}

Model Pricing:

• Gemini Flash-Lite: Free tier (current cloud default)
• GLM-4.7-Flash: Free (local via Ollama)
• Haiku 4.5: $0.80 / $4.00 per 1M tokens (paid fallback)

Database Schema

Supabase Tables (Cloud Backend)

Local File Storage (~/.brainlayer/)

Note: embeddings.db belongs to the BrainLayer package, not autonomous.

Deployment Architecture

Security

• 1Password for secrets — never hardcode API keys
• Supabase RLS — row-level security on all tables
• Separate API keys per project — different keys for Golems vs SongScript
• State sync over HTTPS — encrypted Mac ↔ Railway communication
• Event audit log — all actions logged to golem_events table

Next Steps

1. Read Railway Deployment to set up Supabase and Railway
2. Check Golems to understand each domain expert
3. Review Environment Variables for env var reference