← all repositories
liliMozi/openhanako

Your AI coworker with a desk, a memory, and opinions

A desktop-native agent framework that treats "personality" and "safety sandbox" as first-class features, not afterthoughts.

4.6k stars TypeScript AgentsLLMOps · Eval
openhanako
Velocity · 7d
+54
★ / day
Trend
steady
star history

What it does HanaAgent is an Electron-based AI assistant that runs locally, operates your computer, browses the web, manages files, and chats across Telegram, WeChat, QQ, and Feishu. It wraps a Node.js Hono server with a React frontend, using the Pi SDK for agent orchestration and better-sqlite3 for state. Multiple agents coexist, each with isolated memory, personality files, and a “desk” workspace they monitor for tasks.

The interesting bit The author, a self-described “clerk by day,” built this specifically to escape the command-line tyranny of tools like Claude Code. The result is unusually clerk-friendly: agents import/export as zip “character cards,” skills drag-and-drop into bundles, and a full OS-level sandbox (Seatbelt on macOS, Bubblewrap on Linux, restricted tokens on Windows) keeps the agent from wandering where it shouldn’t. The Windows version notably lacks code signing and runs in a weaker “write-isolation” sandbox—honest rough edges, not buried footnotes.

Key highlights

  • Multi-agent with group chat: Agents share a “channel” for collaboration or delegate tasks to each other.
  • Skill autonomy: Agents can fetch community skills from GitHub or write their own, with strict opt-out sandboxing for installs.
  • Cross-platform bridge: Same agent, same session history, whether you’re on desktop, mobile PWA, or messaging it from WeChat.
  • SessionFile sidecar: Files uploaded anywhere (desktop, bridge, mobile) get a unified identity across all frontends.
  • Plugin architecture with teeth: Two-tier permission model; full-access plugins can inject HTTP routes, LLM providers, and UI widgets.

Caveats

  • Memory system is self-described as “solid for recent events” but explicitly flagged as needing work.
  • Windows is beta, unsigned, and SmartScreen will complain on first run.
  • Anthropic OAuth is omitted due to “ban risk”; OpenAI OAuth is available instead.

Verdict Worth a look if you want a local, multi-personality agent with actual GUI affordances and granular safety controls—not another terminal wrapper. Skip if you need mature Windows hardening or a memory system you can trust for long-horizon recall today.

heatdrop uses Google Analytics to see which pages get read — nothing else. Your call. How we handle data.