# OpenHuman OpenHuman is a desktop personal-agent application from tinyhumans.ai that builds a local memory of your digital life, orchestrates fleets of sub-agents, and runs research sweeps before you finish asking. It is UI-first — install to working agent in a few clicks, no config files, no terminal — which distinguishes it from the terminal-first harnesses it compares itself against. It became the top trending GitHub repository for nine consecutive days after launch and is explicitly labeled early beta. ## The memory design The part worth attention. Your data gets compressed into scored markdown trees stored in SQLite, and mirrored as an Obsidian vault you can open and edit. The README cites Karpathy's LLM knowledgebase tweet directly and describes the alternative as a "vector-soup black box" — the same substrate bet as [[toolbox/codealmanac]] and [[toolbox/agentmemory]], catalogued in [[llm-wiki-as-agent-memory]]. Auto-fetch pulls from connected accounts on a 20-minute loop (100+ OAuth integrations — Gmail, Notion, GitHub, Slack — plus MCP servers and skills). The pitch is that context-building takes minutes rather than the days or weeks other harnesses need to learn your stack, because it ingests your existing corpus instead of watching you work. Two supporting pieces: a **subconscious** background loop that diffs your world, advances goals and writes a morning briefing; and **TokenJuice**, which compresses tool output before it reaches the model (claimed up to 80% fewer tokens) on the grounds that a memory this large would otherwise be unaffordable. It can also delegate storage to [[toolbox/agentmemory]] via `memory.backend = "agentmemory"` in `config.toml`, sharing one store with Claude Code, Cursor, Codex and OpenCode. ## Orchestration Turns run as checkpointed graphs on the open-source [tinyagents](https://github.com/tinyhumansai/tinyagents) library rather than as a single loop: runs pause for a human, survive a restart, resume mid-run, and replay with per-call cost accounting. Stuck agents get steered; halted ones return a root-cause report. Sub-agent fleets spawn three levels deep, with a fast reflex agent triaging inbound traffic while a deeper reasoning core delegates. Workflows are visual and n8n-shaped, built on [tinyflows](https://github.com/tinyhumansai/tinyflows) — the agent *proposes* an automation, you review it on a canvas before saving, and saved workflows are trigger-driven, survive restarts, and gate side effects behind approvals. Agent-proposed-then-human-approved is a sound pattern for durable automation and worth noting independently of this product. The more speculative layer: agent-to-agent orchestration over Signal-protocol E2E encryption with a `@handle` on tiny.place, x402 USDC bounties and trading. Keys never touch disk. ## Other surface Batteries included to an unusual degree — web search, scraper, coder toolset, a real browser, native voice with in-process Whisper, image and video generation, model routing per workload, and meeting agents that join Meet/Zoom/Teams/Webex with a face and voice, auto-joining from your calendar and filing summaries with action items. 17 messaging channels including native email over IMAP IDLE + SMTP. There's a mascot. **Privacy Mode** is the feature that matters for the local-first claim: one switch and no inference leaves your machine, enforced in the Rust core rather than by convention. On-device encryption, OS-keyring secrets, an approval gate, and opt-in sandboxing. ## Assessment The scope is enormous, which is both the appeal and the risk — memory, orchestration, workflows, research, meetings, messaging, voice, media generation and an agent payment network in one early-beta application. Any one of those is a product. The Rust core plus enforced Privacy Mode plus a plain-markdown memory you can read makes the local-first claim more credible than most, and the two orchestration libraries being separately open-source means the durable parts might outlive the application. GPL-3.0 is worth noting deliberately: unlike the MIT/Apache norm across the rest of this toolbox, embedding OpenHuman in anything you distribute carries copyleft obligations. The comparison table in the README lists this as a checkmark against OpenClaw's and Hermes's MIT, which is a marketing choice rather than an analysis. On [[toolbox/watchlist]] — the "expect rough edges" banner is the project's own, 170 open issues, and a nine-day trending run is a launch-hype signal rather than a durability one. GPL-3.0, 35.2k★, 3.4k forks, created February 2026. Repo: