OpenHuman (Tiny Humans AI)
An open-source (GPL-3.0, ~30.7k★) local-first personal AI agent — a Tauri desktop app (Rust 62% / TypeScript 35%) that is a fourth entrant on the self-driven autonomous-assistant pole of claude-code-channels-vs-openclaw, explicitly positioning against claude-cowork, OpenClaw, and hermes-agent. Its pitch: get an agent “context in minutes, not weeks.”
What distinguishes it — local-first + a memory tree
- Memory Tree + Obsidian Wiki — a local-first knowledge base in SQLite on your machine, “inspired by Karpathy’s obsidian-wiki workflow.” This is the strongest tie to the tools-for-thought lineage: it’s a gbrain-style (cross-wiki) compounding personal knowledge base as the agent’s memory layer, and the reason this source’s runner-up home is research-wiki.
- Privacy / local-first — runs on-device, optional Ollama for local models; GPL-3.0 (a copyleft contrast to the permissive licenses common elsewhere here).
- TokenJuice — a context-compression layer claiming up to 80% reduction (a context- engineering move, cf. the progressive-disclosure savings the wiki tracks).
Harness capabilities
- 118+ third-party integrations via OAuth with auto-fetch every 20 min (Gmail, Notion, GitHub, Slack, Stripe, Linear, Jira…), through a Composio connector/tool-proxy layer.
- Full coder toolset (filesystem, git, lint, test, grep) — so it’s also an agentic-coding-harness, not only a chat assistant.
- Model routing for automatic LLM selection; web search/scraping; voice (STT in, ElevenLabs TTS out — cf. claude-to-speech); desktop mascot that joins meetings.
Where it sits
It advances the self-driven pole alongside hermes-agent (open-source successor) and microsoft-scout (enterprise product), but on a distinct axis: local-first + privacy + a persistent markdown/SQLite memory tree, where Hermes is cloud/multi-platform and Scout is Microsoft-365 cloud. Persistent personal memory makes it a durable-agents instance; whether it authors its own skills (self-improving-agents) isn’t claimed — the auto-fetch/memory loop is accumulation, not skill self-editing. Early Beta. (Caveat: README/landing claims, unbenchmarked.)
Related
agentic-coding-harness · claude-code-channels-vs-openclaw · hermes-agent · microsoft-scout · claude-cowork · durable-agents · gbrain · claude-to-speech