// the find
tinyhumansai/openhuman
Your Personal AI super intelligence. Private, Simple and extremely powerful.
OpenHuman is a Tauri-based desktop AI assistant that combines 118+ OAuth integrations, a local SQLite memory tree, a desktop mascot, and voice capabilities into a single app. It targets people who want an AI agent with persistent context across their email, calendar, Slack, GitHub, etc., without manually feeding it data. The star count is high but it's early beta with the usual rough edges that implies.
- The Cargo.toml comments are unusually honest and informative — there's a real profiling note explaining why html2md was dropped (894MB peak heap on 10KB HTML), and comments explaining why specific crate versions and features were chosen. This is the sign of engineers actually debugging production problems.
- Architecture is well-structured: Rust core for business logic/RPC server, Tauri shell for desktop, separate scanners per messaging platform (WhatsApp, Telegram, Slack, Discord each get their own module with DOM snapshot, IDB, and extraction layers), and TypeScript frontend. Separation of concerns is clear.
- TokenJuice token compression as a first-class feature rather than an afterthought is genuinely useful — converting HTML to Markdown, deduping tool output, and grapheme-preserving CJK handling before touching any LLM is the kind of thing most agent tools ignore until cost becomes a problem.
- The dev setup, while heavy (Node 24+, pnpm, Rust 1.93, CMake, Ninja, ripgrep, CEF submodules), is well-documented with explicit versions, a beginner contributor guide, and multiple focused CI workflows covering typecheck, coverage, e2e per integration, and installer smoke tests.
- The 'privacy-first' claim sits uncomfortably next to the managed integrations path being proxied through OpenHuman's Composio connector layer. Your OAuth tokens and integration data route through their backend by default, and self-hosting requires bringing your own Composio key plus hosting your own webhooks. That's a significant asterisk on the privacy story.
- Build complexity is brutal for contributors: vendored Tauri/CEF submodules, a custom tauri-cef fork, CMake/Ninja requirements, and a CEF SHA pin guard workflow. Anyone not on a well-specced dev machine will spend hours just getting to `pnpm dev`, and the Windows build is a separate CI workflow suggesting it's a second-class path.
- The comparison table references 'Claude Cowork', 'OpenClaw', and 'Hermes Agent' — none of which appear to be real or widely-known products. This reads as marketing fiction designed to make OpenHuman's feature matrix look favorable, which undermines trust in the rest of the claims.
- Early beta + 24k stars means a lot of people will install this expecting polish they won't find. The Google Meet integration injects JS and uses a fake camera/audio bridge, which is the kind of thing that breaks silently whenever Google updates Meet's DOM, and there's no obvious signal in the README about how stable any individual integration actually is.