Loading...
Golems is a Bun monorepo of AI agent tooling, published in the open as my working setup. Domain packages each own one area: coach handles scheduling and calendar integration, teller tracks subscriptions and spending, and services runs a morning briefing. golem-powers is a library of 50+ SKILL.md workflows for coding agents like Claude Code, Codex and Cursor, many with executable evals.
58 skills with 3-layer architecture: SKILL.md (universal) + adapters/ (per-CLI) + capabilities.yaml (routing).
37 skills ship eval suites with structured assertions and fixture-based testing. No results are published until runs on current models record the model used.
Telegram
User commands
OrcClaude v2
Multi-agent sprints
58 Skills
4 CLI adapters
Eval Framework
37 eval suites
Cloud + Local
Railway + Mac
Telegram
User commands
OrcClaude v2
Multi-agent sprints
58 Skills
4 CLI adapters
Eval Framework
37 eval suites
Cloud + Local
Railway + Mac
git clone https://github.com/EtanHey/golems.git
cd golems && bun installStarted as a single-file Telegram bot. Grew into an autonomous agent ecosystem. Componentized through a 9-phase plan into 10 Bun workspace packages — each a self-contained Claude Code plugin.
Each golem owns a domain: Claude (orchestrator + Telegram), Jobs (scraping + matching), Recruiter (outreach + practice), Teller (finance), Content (visual + text), Coach (calendar + planning), Services (Night Shift + briefing + cloud worker).
Mac = Brain (Telegram bot, Night Shift, local LLMs). Railway = Body (email poller, job scraper, morning briefing). Multi-LLM routing: Gemini Flash-Lite (free cloud) → GLM-4.7-Flash (free local) → Haiku (paid fallback).
8 MCP servers exposing 60+ tools: BrainLayer (memory), VoiceLayer (voice I/O), email triage, job discovery, GLM summarization, Supabase access, web search, and browser automation.
Ralph executes PRD stories autonomously: read story → implement → CodeRabbit review → fix → commit. Night Shift runs at 4am scanning repos for TODOs, creating worktrees, and shipping improvements while the human sleeps.
BrainLayer gives agents persistent memory. VoiceLayer gives them voice. Together with Golems orchestration: agents that remember, speak, listen, and act autonomously across every project.
Persistent memory for AI agents over MCP. 16 tools, hybrid vector + full-text search and a local knowledge graph over one SQLite file. Bounded TLA+ models checked by TLC. pip install brainlayer.
2 MCP tools (voice_speak + voice_ask) + 9 QA tools, whisper.cpp STT (~300ms), edge-tts, VoiceBar MCP daemon, session booking. 546 tests. bunx voicelayer-mcp.
Orchestrator agent with planning topology, response markers, and multi-agent sprint coordination. Spawns and monitors parallel Claude workers.
Night Shift runs improvements at 4am with CodeRabbit review gates. PR loop v2 enforces review on every commit. Morning Briefing at 8am.
Railway for cloud worker (email, jobs, briefings). Mac for Telegram bot, memory, voice. ~$5/month total.