What is Miru?

Miru (見る) is an open-source AI companion designed to run entirely on your own hardware — macOS, Windows, Android, or a self-hosted Linux server. Unlike typical chatbots, Miru lives as a Live2D desktop pet that quietly observes your screen (with permission), builds verifiable long-term memory in plain Markdown (projects, people, yourself, topics, daily journal, and commitments), and proactively reaches out via an Attention Engine that reads your context and mood. It supports multi-device sync through an invitation code connecting clients to your own server. You provide your own OpenAI-compatible model API keys for three tiers (Vision, Chat, Memory). The software is free under Apache 2.0; Live2D components remain under their own licenses. No central cloud or telemetry — all data stays on your machines.

Should you self-host Miru?

Miru scores 43/100 on selfhost.directory (how scoring works) and ranks #78 of 120 in AI Assistant. It has 187 GitHub stars.

It shipped 2 stable releases in the past year. The latest, cross-platform-2026-08-30, came out on 30 August 2026. New versions land roughly every 5 weeks.

Setup is advanced: plan for several services and some ongoing maintenance. It is Apache-2.0-licensed and written mostly in Python.

Its closest alternative on selfhost.directory is hermes-agent (89/100, 257 885 stars). It scores higher overall, so compare the two before you pick. See Miru vs hermes-agent side by side.

Features

Commitment tracking
Emotion and personality system
Live2D desktop pet
Long-term verifiable memory
Multi-device sync via self-hosted server
Proactive Attention Engine
Screen sensor with privacy
Sleep agent for nightly memory organization
SSE real-time sync across devices
Three-tier model architecture

FAQ

她和普通 AI 聊天到底不一样在哪？

Unlike regular chat that forgets each conversation, Miru has continuous memory and emotion, proactively reaches out, and runs entirely on your own hardware — you can see everything she remembers and how she thought.

她说「记得」，我能亲眼验证吗？

Yes. The app has a memory panel with four card categories (projects, people, yourself, topics), daily journal, and commitment list — all openable and exportable as Markdown. Memory writes only come from real events.

用起来要花多少钱？

Software is free. You provide your own three model API keys (Vision/Chat/Memory) for any OpenAI-compatible service. Official recommended config costs under ¥2/day in tokens.

她要看我的屏幕，安全吗？

Default every 60 seconds, skips if screen unchanged; frequency adjustable or can be disabled entirely. Screenshots discarded after analysis; only understanding kept. All processing happens on your own machine.

她会一直主动弹消息打扰我吗？

No. She reads your screen and rhythm to judge the moment — stays quiet during deep work, reaches out when you're stuck, tired, or relaxing. It's her decision, not a timer or notification rule.

我需要准备什么？手机也能用吗？

Two modes: local single-device (Mac only, DMG) or multi-device via self-hosted Linux server. Android works now, iOS is on the roadmap. No central cloud — all your own machines.