openhuman

OpenHuman is an open-source personal AI application for macOS, Windows and Linux, currently in early beta. It is positioned as three things: a persistent local memory system, an agent orchestrator, and a deep research assistant. The project is built with a Rust core and a desktop shell, and is distributed through installers on the vendor site and GitHub Releases, plus Homebrew, Debian/Ubuntu packages, AUR and install scripts. Memory and context The core memory feature is the Memory Tree: user data is compressed into scored Markdown trees stored in SQLite on the local machine, and mirrored as an Obsidian vault that can be opened and edited directly. An auto-fetch loop pulls data from connected accounts roughly every 20 minutes so the agent has recent context without a long training period. An optional agentmemory backend can be configured in config.toml so the same durable store is shared with other coding agents. Integrations and tools The README claims 100+ OAuth integrations, 5,000+ MCP servers and 90,000+ Skills, covering services such as Gmail, Notion, GitHub and Slack. Native tools include managed web search powered by Exa (included with a subscription, or bring-your-own key), a scraper, a coder toolset, a browser and computer-use tool, and native voice with in-process Whisper. Image and video generation are also listed. Model routing selects an LLM per workload; users can instead point workloads at their own provider keys or a fully local Ollama model. Orchestration and workflows Rather than a single agent loop, OpenHuman runs turns as checkpointed graphs on the open-source tinyagents library, so runs can pause for human input, survive restarts and resume. Sub-agents can spawn several levels deep, and stuck agents are reported with root causes. Agent-to-agent messaging is described as running over Signal-protocol end-to-end encryption with x402 payments. Workflows, built on the open-source tinyflows project, are proposed by the agent and reviewed by the user on a visual canvas before being saved; saved workflows are trigger-driven (schedules, webhooks, channel events) and gate side effects behind approvals. A TokenJuice component compresses tool output before it reaches the model. Channels and privacy Seventeen messaging channels are listed, including Telegram, Discord, Slack, WhatsApp, Signal and iMessage, plus native email over IMAP IDLE and SMTP. Privacy features include on-device encrypted data, an approval gate, OS-keyring secret storage, opt-in sandboxing, and a Privacy Mode switch that prevents inference from leaving the machine, enforced in the Rust core. Theming includes several theme families and a visual editor with JSON export. Development Contributors need Git, Node.js 24+, pnpm 10.10.0, Rust 1.93.0 with rustfmt and clippy, CMake, Ninja, ripgrep and platform desktop build prerequisites. The workflow involves forking and cloning, initializing submodules before installing dependencies, and using pnpm scripts for web-only UI work, the desktop shell, typechecking, formatting and cargo checks. The project is licensed under GNU and links to architecture and setup documentation.