// the find
localgpt-app/localgpt
Local AI assistant, dreaming explorable worlds.
Two tools sharing a repo: a local-first AI assistant (localgpt) and a Bevy-based 3D world builder (localgpt-gen). The core CLI is a genuine single Rust binary with no Node or Python runtime dependency, supporting local models via Ollama/LM Studio as well as cloud providers. It targets developers who want a Claude Desktop alternative that keeps data on their machine.
The security model is real engineering, not a README bullet: Landlock/seccomp on Linux, Seatbelt on macOS, HMAC-signed policy files, and a hash-chained audit log. SQLite FTS5 + sqlite-vec for memory avoids the 'install a separate vector database' problem that kills most local AI projects. MCP server mode plugs straight into Claude Desktop, Cursor, or any MCP-compatible editor without extra plumbing. The localgpt-gen Bevy integration is a genuinely unusual idea — structured LLM tool calls that emit parametric geometry, PBR materials, spatial audio, and entity behaviors into a live 3D scene rather than just generating code.
The project claims to have Android apps, Apple apps, five chat bridges, an MCP server, a 3D world builder, and a full AI assistant — all at 1108 stars, reportedly started in four nights. The Android app is essentially boilerplate scaffolding with a ChatViewModel; the mobile story is not finished. The WhatsApp bridge has an adapter.js sitting in it, which quietly contradicts the 'no Node.js required' claim. The 3D world generation has no quality feedback loop — the LLM either produces the right scene or you prompt again, and there is nothing in the repo that validates or corrects generated output. The markdown-based memory system auto-loads MEMORY.md every session, which scales fine at small sizes but has no documented eviction strategy; as the knowledge store grows, you will hit context limits without warning.