OpenHuman is a local-first, open-source desktop AI agent that TinyHumans beta-launched in May 2026. Built with Rust and Tauri v2, the native binary runs on macOS, Windows, and Linux. It accumulates user's email, messaging, notes, and calendar data in local SQLite and Obsidian-compatible Markdown vaults (~/.openhuman/wiki/) to form persistent memory. Just as Obsidian revolutionized personal knowledge management with Markdown files, OpenHuman aims to be a personal AI runtime that, with long-term memory, deeply understands the user the more it is used. Within 15 days of its release, it garnered over 32,000 stars on GitHub and ranked #1 simultaneously on Product Hunt's daily and weekly charts. Existing AI agents have two structural limitations. Cloud services like ChatGPT and Gemini operate on a session basis, losing context after a conversation ends, and sensitive research data is transmitted to external servers. On the other hand, local LLM solutions like Ollama and LM Studio ensure privacy but lack external service integration or systematic memory management capabilities. OpenHuman addresses this gap with its Memory Tree architecture. Source adapters automatically collect data from Gmail, Notion, and Slack every 20 minutes (TICKSECONDS=1200), normalize it into chunks of 3,000 tokens or less, and organize it into three tree structures: source-based, topic-based, and daily digest. Each leaf node goes through a lifecycle of pendingextraction, admitted, buffered, and sealed, and is promoted or discarded based on its deep score. The TokenJuice compression engine reduces the tokens of tool output by approximately 70%, significantly lowering API costs, and the model routing system automatically selects the optimal model from over 30 LLM providers based on task type hints (e.g., hint:reasoning, hint:fast, hint:vision, hint:code). It supports one-click OAuth integration with over 118 services via Composio connectors, and from v0.54.0, it can share memory with other AI coding tools such as Claude Code, Cursor, and Codex through the agentmemory backend. From the perspective of a life science researcher, OpenHuman can serve as a hub for research workflows. During the process of reviewing papers, journal notifications received via Gmail, reading notes organized in Notion, and discussion content from Slack lab channels are automatically accumulated in the Memory Tree, allowing for immediate and contextual responses to queries such as "Summarize the keywords that have been repeatedly mentioned in CRISPR delivery-related papers over the past month." With its built-in coder toolset (filesystem, git, lint, test, grep), Python analysis scripts can be debugged or pipelines automated using natural language, and voice input (STT via Whisper) and voice output (TTS via Piper) are supported, allowing for interaction with the agent even in wet lab settings where hands are occupied. When local AI mode is activated, the Gemma3 1B model and the all-minilm embedding model perform summarization and search even in offline environments, ensuring complete privacy even in highly secure clinical data environments.

Local-first, open-source desktop AI agent — designed with privacy in mind.

Research Literature Knowledge Hub: Gmail journal notifications, Notion reading notes, and Slack lab discussions are automatically aggregated into Memory Tree every 20 minutes. It generates context-aware responses with source- and time-based weighting for queries like, "Summarize the tools that were repeatedly mentioned in the scRNA-seq related discussions this quarter." This saves several hours per week compared to manual literature organization.