What it does OpenHuman is a Rust-built desktop agentic assistant that runs locally and plugs into your existing tools — Gmail, Slack, Notion, GitHub, Calendar, and 110+ others. It maintains a local SQLite knowledge base (the “Memory Tree”) and an Obsidian-compatible Markdown vault, auto-fetching fresh data every twenty minutes so the agent has context without you copy-pasting.

The interesting bit The project treats “local-first” as infrastructure, not ideology: your data sits in SQLite and .md files you can browse, but model routing, web search, and OAuth handshakes default through OpenHuman’s managed backend. You can bring your own LLM via Ollama and wire your own Composio credentials, but some real-time triggers still need their servers. It’s a pragmatic hybrid — more “self-hosted with escape hatches” than pure offline fantasy.

Key highlights

• Desktop mascot with voice (STT in, ElevenLabs TTS out) that can join Google Meet calls as a participant
• TokenJuice compression layer: HTML→Markdown, deduping, summarization before any LLM sees it; claims up to 80% token reduction
• 118+ integrations via Composio with one-click OAuth, exposed as typed tools to the agent
• Native packages with proper signing chains (Homebrew, signed apt repo, MSI) — unusually careful for an early-beta project
• Optional local AI via Ollama for supported workloads

Caveats

• Early beta with acknowledged rough edges; AppImage crashes on Wayland and some Arch setups (issue #2463)
• Default managed mode requires OpenHuman-hosted services for sign-in, model routing, and search proxying; full self-hosting isn’t there yet
• Script installs (curl | bash) are explicitly unverified — the README warns against them and prefers native package paths

Verdict Worth a look if you want an agent that remembers your context across weeks and doesn’t live in a chat tab. Skip it if you need fully offline operation or are allergic to managed-service defaults.