Open-source personal AI assistant that runs on your own machine and answers from WhatsApp, Telegram, Slack, iMessage and 25 other chat apps.
Summary
Open-source desktop AI agent that builds a local, readable memory of your work, then orchestrates agent fleets and durable workflows on top of it.
Screenshots
Description
OpenHuman is an open-source desktop agent that inverts the usual order of operations: instead of starting each conversation from nothing, it spends its idle time building a local, readable model of your work — your mail, calendar, documents, repositories and tickets — so that the first prompt of a session already has context behind it. Built by TinyHumans AI, it reached the top of GitHub's trending list shortly after launch and now carries north of 36,000 stars.
A memory you can read
The memory layer is a Memory Tree stored as compressed Markdown in a local SQLite database, browsable as an Obsidian wiki. That design choice matters more than it sounds: because memory is plain text on your disk rather than opaque embeddings in someone else's vector store, you can audit what the agent believes about you and delete what you would rather it forget. Connected sources re-sync on a roughly 20-minute cycle, and a background loop keeps working on open threads after you stop typing.
Orchestration, not just chat
Runs are checkpointed graphs on the project's own tinyagents runtime, so a long task survives a crash rather than restarting. Sub-agent fleets can be nested, with recovery for agents that get stuck, and tinyflows handles durable trigger-driven workflows — the agent can propose an automation and hand it to you on a canvas for review before it becomes standing behaviour. Agent-to-agent sessions are encrypted with the Signal protocol.
Reach
Over 100 OAuth integrations cover the usual suspects — Gmail, Notion, GitHub, Slack, Stripe, Google Calendar and Drive, Linear, Jira — alongside MCP server and Agent Skills support, native Whisper voice, browser automation, meeting capture for Meet, Zoom, Teams and Webex, and 17 messaging channels including direct IMAP/SMTP.
Privacy posture and platforms
Data is encrypted on device, secrets sit in the OS keyring, and sensitive operations are approval-gated. A Privacy Mode enforced in the Rust core restricts inference to local models for users who want nothing leaving the machine. It ships as a Tauri v2 desktop app for macOS, Windows and Linux — installers, Homebrew, a Debian package, the AUR, or Docker — under GPL-3.0. The project labels itself early beta and releases most days.
Reviews
Similar App Suggestions
Native macOS harness for AI agents — any model, persistent memory, sandboxed execution and cryptographic identity, fully offline and open source.
Andrew Ng's open-source desktop AI coworker that delivers finished work, not just chat replies.
Open-source desktop app and MCP for running AI agents on your own files, an open alternative to Claude Cowork.
Moonshot AI's assistant, known for very long context windows and strong agentic coding and research workflows.