About this project

LocalGPT is an open-source, local-first AI assistant written in Rust, distributed as a single binary with no Node.js, Docker, or Python requirement. It ships two main components: an AI assistant and a world-building tool. AI Assistant (localgpt) - Interfaces: interactive CLI chat, single-question mode, and a daemon exposing an embedded web UI and HTTP API (chat, SSE streaming chat, memory search). - Persistent memory: a markdown-based workspace with MEMORY.md, HEARTBEAT.md (autonomous task queue), SOUL.md (personality guidance), and a knowledge/ bank. Files are indexed with SQLite FTS5 for keyword search and sqlite-vec for semantic search using local embeddings. - Autonomous heartbeat: tasks can be delegated and run in the background. - Providers: LM Studio, Ollama, Anthropic, OpenAI, xAI, GLM, Vertex AI, and CLI providers; configuration lives in config.toml with XDG-compliant paths. - Security: kernel-enforced sandboxing (Landlock/seccomp on Linux, Seatbelt on macOS), a protected POLICY.md with standing instructions, and prompt-injection defenses such as marker stripping and content boundaries. Gen Mode (localgpt-gen) - Builds explorable 3D worlds from natural-language prompts on the Bevy engine. - Supports parametric shapes (box, sphere, cylinder, capsule, plane, torus, pyramid, tetrahedron, icosahedron, wedge), PBR materials, point/spot/directional lighting, behaviors (orbit, spin, bob, look_at, pulse, path_follow, bounce), and ambient or spatial audio. - Exports glTF/GLB, browser-viewable HTML, and screenshots; worlds can be saved and loaded as reusable skills. - Headless mode and an experiment queue allow batch generation without a window, useful for overnight runs or CI pipelines; a memory system learns style preferences across sessions. - An MCP server lets Gen be used from Claude Desktop, Codex Desktop, and MCP-compatible editors such as VS Code, Zed, and Cursor, with a CLI relay mode for command-line tools. Installation is via crates.io (cargo install localgpt-gen, cargo install localgpt) or from source with cargo run. The project is licensed Apache-2.0 and built with Rust, Tokio, Axum, Bevy, SQLite (FTS5 + sqlite-vec), fastembed, and eframe.