Out-of-the-Box · Open-Source · Universal · Vendor-Neutral

English | Català | Čeština | Dansk | Deutsch | Ελληνικά | English (UK) | Español | Français | Gaeilge | Hrvatski | Magyar | Italiano | 日本語 | 한국어 | മലയാളം | Norsk Bokmål | Nederlands | Polski | Português (BR) | Português (PT) | Română | Русский | Slovenčina | Svenska | 简体中文 | 繁體中文

Inspired by OpenClaw, we believe the future of personal computing will be shaped by diverse, local-first AI agents running at the edge.

ZimaOS Blue is our answer — a fully open-source, auditable, vendor-neutral, and production-ready agent runtime and toolkit that lets you ship private, self-hosted agents with zero friction.

Built for bold developers who want to vibe or handcraft their own agents, Blue is engineered for performance: written in Go, with a memory footprint as low as 19 MB. It runs on any x86, ZimaOS, Raspberry Pi, Windows, macOS — anywhere you plug in power.

A quick demo of conversation flow and task execution in Blue.

A quick demo of Blue's LLM providers integration experience.

A quick demo covering the overall product overview, channels, and additional configuration.

100% Go, static binary. Cross-compiles to 5 targets out of the box ( linux/amd64, linux/arm64, darwin/amd64, darwin/arm64, windows/amd64). No Node runtime, no Python, no containers required. Drop it on a NAS, a ZimaOS, a Raspberry Pi, an old x86 router, or a Mac — it just runs. Then layer on your own UI, logic, and agent skills — one codebase, every platform.

Everyone wants tools that are simple, reliable, and scale when you need them. Tools that just work, so you can focus on what you're actually building.

This isn't a new philosophy. It's the same one that built ZimaOS: simple, reliable, and built to stay out of your way. Blue is that philosophy, extended to the agent stack.

From deep research that delivers a full HTML report, to OCR, PDF, browser automation, and document conversion, Blue handles complex, real-world workflows without sending your data to the cloud. Voice wake, STT/TTS, Talk Mode, and support for local inference make everyday interactions instant, private, and always available.

Get the native application — no dependencies, no compilation. Built-in trial configuration with onboarding in seconds — start chatting instantly via remote connection, no bot setup required. True out-of-the-box experience.

• macOS: Download DMG
• Windows: Download Installer

ZimaOS / macOS / linux

curl -fsSL https://ota.zimaos.com/blue | sh

Windows (PowerShell)

irm https://ota.zimaos.com/blue/windows | iex

git clone https://github.com/IceWhaleTech/ZimaOS-Blue.git
cd ZimaOS-Blue
git submodule update --init --recursive

ZimaOS / macOS / linux

sh build.sh

Windows (PowerShell)

.\build.bat

Note: Windows builds require:

• MinGW-w64 (gcc) and CMake for native C dependencies (espeak-ng, whisper.cpp, opus, kokoro, onnx)
• Windows SDK for system libraries (winmm, etc.)

Make sure gcc, cmake are in your PATH.

Take it further: it delivers native support for 20+ IM platforms, voice‑driven interfaces for natural, context‑aware dialogue, zero‑config model switching with IDE scanning.

⚠️ [!IMPORTANT]

If you plan to keep tuning or vibe coding on top of Blue, do not treat a few good-looking chats as release evidence. Any change that affects routing, execution behavior, tool surface, budget control, model selection, or the execution framework should be validated with Blue Harness, not with ad hoc spot checks.

Blue should follow one simple rule here: data first, gates first, cut over last. In practice, that means updating the relevant Harness dataset / eval spec before judging a change, then keeping one stable candidate_id across the whole attempt so selector, execution, budget, and readiness reports all describe the same candidate instead of four unrelated runs.

1. Run blue harness selector verify
2. Run blue harness execution verify
3. Reuse the selector eval run for blue harness budget gate
4. Finish with blue harness cutover-readiness

For local iteration, nightly validation, or CI evidence collection, prefer python3 scripts/cutover_candidate_pipeline.py. It runs the full selector -> execution -> budget -> readiness sequence under one shared candidate, which makes the result easier to compare, review, and cut over from.

In short, tuning on top of Blue is not about "it feels better in a few chats." It is about putting the candidate into Harness, collecting comparable evidence, and letting the gate and readiness results decide whether the change is truly safe to keep.

• Issues: Please file bugs and feature requests here
• Discussions: Discord
• Follow us on GitHub

This project is licensed under the MIT License - see the LICENSE file for details. We believe in open source and giving back to the community.

Thanks to all Blue contributors:

1. OpenClaw — Local-first open source agent. Pioneered connecting LLMs to local devices through channel adapters and tool calling, directly inspiring Blue's agent runtime architecture. https://github.com/openclaw/openclaw
2. MiroMind — Deep research mode with evidence-backed synthesis. Shaped Blue's built-in deep research pipeline: planning, parallel retrieval, evidence deduplication, and HTML report generation. https://www.miromind.ai
3. Karpathy's LLM Wiki — LLM as knowledge compiler. Reframes LLMs to build persistent, evolving knowledge spaces, moving beyond RAG's accumulation trap.
4. OpenSpace (HKUDS) — Self-evolving skill engine. A DAG-based framework where agents learn from failures and derive specialized skills. https://github.com/HKUDS/OpenSpace
5. Andrew Ng's Context Hub — Versioned API documentation registry for coding agents. Addresses agent hallucinations and forgotten session knowledge. Provides curated, versioned docs with annotation and feedback loops, turning documentation into a self-improving knowledge layer. https://github.com/andrewyng/context-hub
6. Notion — Simple, human, and intentionally quiet. Inspired by the minimalist ethos of Notion, Blue brings warmth back to the grid. Where refined serif meets thoughtful design, crafting a space that feels like home. https://www.notion.com/about
7. Matrix — Visual inspiration from the iconic digital rain aesthetic. The aesthetic direction for Blue's technical diagrams.
8. IceWhale — Love, Death & Robots S2E2 "Ice". A collective that gathers worldwide to break through internet giants' walls and resist data concentration. The ice whale symbolizes a community building sovereign tools together at the edge.
9. ZimaOS Blue — Love, Death & Robots S1E14 "Zima Blue". A metaphor: intelligence that begins in service and evolves to explore the world. Blue is an agent for wisdom, rooted in simplicity and reaching for depth.
10. ZimaOS — Simplified, Focused, Open design principles. Both ZimaOS and Blue share the belief that technology should serve the user — deploy in 30 seconds, run anywhere, stay vendor-neutral. https://www.zimaspace.com/zimaos