Loading...
What Golems - Autonomous AI Agents can do
Same skills, any CLI — Claude, Codex, Cursor, Gemini, Kiro
Skills are written once in universal SKILL.md format, then adapted for each AI CLI via a thin adapters/ layer. A capabilities.yaml file routes each skill to the right adapters based on what each CLI supports. 40 skill eval packs with 480+ assertions ensure quality. 96% pass rate across the eval suite. The adapter layer means skills work across 5 different AI CLIs without rewriting.
PRD stories to working code, unattended
Ralph reads structured PRD stories and spawns fresh Claude instances to implement each one. Every commit is gated behind CodeRabbit AI review — if issues are found, Ralph fixes them automatically (up to 3 attempts). Failed fixes create BUG stories instead of shipping broken code. The cycle continues until all stories are complete.
{
"id": "US-001",
"title": "Add session export",
"criteria": [
"Export sessions to JSON format",
"Include all enrichment metadata",
"Run CodeRabbit review - must pass",
"Commit: feat: US-001 add session export"
]
}PRD story — last 2 criteria are always CodeRabbit + commit
Multi-agent orchestrator with planning topology
The orchestrator agent coordinates multi-agent sprints across repos. Planning topology with response markers enables structured delegation. Spawns parallel Claude workers, monitors progress via collab files, and dispatches research to specialized agents. Sequential-parallel collab chains enable fully automated handoffs between agents — zero human intervention.
Night Shift + PR Loop v2 — every commit reviewed
Night Shift scans repos at 4am for TODOs and improvements, creates worktrees, implements changes, and gates every commit behind CodeRabbit AI review. PR Loop v2 enforces review on every commit — if issues are found, they're fixed automatically (up to 3 attempts). Failed fixes create BUG stories instead of shipping broken code. Ralph reads PRD stories and spawns fresh Claude instances to implement each one.
Railway for cron, Mac for real-time
Railway hosts a single cloud worker running scheduled tasks: email polling (hourly), job scraping (3x/day), and daily briefing generation — all using free Gemini Flash-Lite. macOS handles everything real-time: the Telegram bot on port 3847, BrainLayer memory indexing, VoiceLayer for voice I/O, and Night Shift coding. Total cloud cost: ~$5/month.
8 MCP servers powering every golem
Each golem declares which MCP servers it needs. BrainLayer provides 12 memory + KG tools including the new brain_digest with 3 modes and pubsub for real-time updates. VoiceLayer exposes 2 voice tools with daemon architecture. The email server handles triage with 7 tools. Plus Supabase for database, Exa for web search, Sophtron for financial data, and GLM for local inference.
2D canvas knowledge graph + enrichment explorer
A Next.js dashboard with d3-force 2D canvas knowledge graph visualization, enrichment observatory for search quality analysis, and wiki synthesis panels. Entity detail panels with community clustering let you explore the knowledge graph interactively. Filter panels handle 284K+ chunks without crashing.