Inspired by the Gods from the Machine raid in Star Wars: The Old Republic — a pantheon of ancient, sentient super-weapons worshiped as deities on the mechanical Dyson sphere planet Iokath.
We build autonomous, AI-powered software — "gods in the machine" — that operate entirely on local, private, offline AI models. No cloud dependencies. No API keys to external services. Just raw local inference powering intelligent programs.
Every project is designed to run on consumer hardware using small, capable models (Qwen 3.5 family: 0.8B, 2B, 4B, 9B) served via llama.cpp.
| Project | Language | Description |
|---|---|---|
| gilgamesh | Go | TDD-driven local AI coding agent (v0.6). CLI + MCP + HTTP API. 7 tools, 7 built-in skills, streaming markdown rendering, graceful Ctrl+C, error classification, config validation, env var overrides, memory persistence, conversation history, 243 tests. Includes Go benchmark suite for model trialing. |
| zeus | Zig | Zig-as-a-build-system for llama.cpp-powered local GGUF inference. (incomplete) |
| raijin | Rust | ONNX CPU inference engine. (incomplete) |
| garuda | Zig | Directory tree viewer. (incomplete) |
| godsfromthemachine.github.io | Hugo | Official project website — architecture, roadmap, and documentation. |
- Local-first: All intelligence comes from free, local, private AI models
- Blazing fast: Built in Go, Rust, Zig, and other high-performance languages
- Lean: Minimal token overhead for CPU inference — first response in seconds, not minutes
- Open source: MIT/GNU licensed — free to use, learn from, and contribute to
- CLI + MCP + API: Every god has CLI, MCP, and HTTP/API interfaces
- Mythological: Projects named after gods and legends — each one a specialized autonomous entity
Every god exposes the same capabilities through three interfaces — because every MCP tool functions just as well as a CLI command given to an agent via a shell, and just as well as an HTTP endpoint called by another program. We build all three so gods can interface with human users, other agents, external programs, and each other.
We run controlled model trials on consumer hardware — no GPU, just CPU inference. Our benchmark suite tests models across 6 stages from raw inference to full agent edit tasks.
Key findings:
- Qwen3.5 2B Q4_K_M: speed sweet spot (19 tok/s TG on CPU)
- Qwen3.5 4B Q4_K_M: quality ceiling (2 tool calls vs 9 for same task)
- 12 threads optimal on 16-core EPYC (16 threads degrades TG by 30%)
- KV cache q4_0 quantization saves 5-7% RAM with no quality loss
- Token budget under 1,600 tokens — competitors use 10,000-40,000
See the full research page and trial methodology.
All projects are open source. Browse the repos, open issues, submit PRs. We're building the future of autonomous local AI software.
godsfromthemachine.github.io — Visit the website for architecture docs, roadmap, and more.
