OpenHuman
A local-first, Rust-based agent harness with one core that runs as a desktop app, browser UI, terminal client and library, connecting to whichever LLM, memory or search engine a user already runs.
OpenHuman's core runs in-process rather than as a separate daemon the interface talks to over a socket. The project's own benchmarks report a marginal memory cost of under 2 MiB per additional agent when many agents share one process, measured with 50, 100 and 500 concurrent agents, against roughly 48 MiB per agent when each runs as its own process. A cold agent turn takes about 102 ms, and a full startup, including config load, registry init, agent build and memory construction, takes about 476 ms.
Cargo feature gates control what compiles into a build, from a roughly 51 MiB stripped minimal build up to a 116 MiB unstripped build with every feature enabled; optional native modules add capabilities such as documents, voice, token compression, a runtime, a wallet, MCP support and messaging channel connectors. The LLM, embedding and memory or search providers it calls are all chosen through configuration, including local options such as Ollama, LM Studio and MLX, Claude Code or the Claude Agent SDK, and more than two dozen bring-your-own-key providers including OpenAI, Anthropic, Google, Groq, Mistral, DeepSeek, Together and Fireworks.
It is written in Rust and licensed under GPL-3.0.