Frequently Asked Questions

RobiClaw is a local-first AI workstation for Windows 11. It runs Gemma 4 language models entirely on your machine using llamafile — Mozilla's portable LLM runtime. A Tauri shell manages the processes and provides a native window with a React UI. No cloud services, no subscriptions, no Docker, no Python install required.

The installer puts three executables on your machine:

• llamafile.exe — the LLM inference engine (port 11434). Exposes an OpenAI-compatible REST API. Runs on CPU or GPU automatically.
• node-api.exe — the Express business-logic server (port 3000). Handles Runners, tools, memory, workspace, and settings.
• RobiClaw.exe — the Tauri shell that spawns the other two as child processes, hosts the React UI in a native webview, and cleans up everything on exit.

All three are bundled inside the NSIS installer. You start RobiClaw.exe and everything else happens automatically.

Minimum (CPU-only, slower):

• Windows 11 64-bit
• 8 GB RAM
• 5 GB free disk (models download separately, ~1.8–3.3 GB each)
• Internet connection for the one-time model download

Recommended (GPU-accelerated):

• Windows 11 64-bit
• 16–32 GB RAM
• NVIDIA RTX with 8 GB+ VRAM (or AMD RX 7000 series)
• SSD with 10 GB free

No Docker, no WSL, no Python, no Node.js needed on the host.

RobiClaw uses Google Gemma 4 GGUF models:

• gemma-4-E4B-it-Q4_K_M.gguf — 3.3 GB, recommended for 8 GB+ RAM
• gemma-4-E2B-it-Q4_K_M.gguf — 1.8 GB, fallback for lower-RAM machines

The onboarding wizard detects your hardware and automatically recommends and downloads the right model from HuggingFace. No account needed.

For semantic memory (optional), it uses nomic-embed-text-v1.5-Q8_0.gguf on a separate llamafile instance (port 11435).

No. All AI inference runs locally via llamafile. Your conversations, files, and memory database stay on your machine. The only external network calls are:

• One-time model download from HuggingFace CDN (during onboarding)
• Optional web search via Brave Search API or DuckDuckGo (Research Runner only, when you explicitly trigger a search)
• Optional GitHub API calls (Project Runner, only if you configure a PAT)

There is no telemetry, no analytics, no account creation, and no licence-check server.

1. Download RobiClaw_1.0.0_x64-setup.exe from GitHub Releases
2. Run the installer (per-user install, no admin rights needed)
3. Launch RobiClaw from the Start Menu shortcut
4. Complete the onboarding wizard (name → hardware check → model download)
5. Start chatting or pick a Runner

The installer is approximately 80 MB. The AI model download (1.8–3.3 GB) happens during onboarding — you can leave it running in the background.

Runners are focused automation modules — think of them as pre-configured AI personas with purpose-built tools and starter prompts. Built-in Runners include:

• 🔬 Research — web search, source synthesis, saved reports
• 💻 Code — read/write files, debug, review, test
• ✍️ Content Creator — blog posts, emails, social threads
• 📄 Document — summarise, draft, extract, format
• 📁 Workspace — find, organise, and manage local files
• 🧠 Memory — store and recall facts across sessions

Custom Runners can be added by creating a RUNNER.md file in ~/Documents/RobiClaw/runners/.

Not yet. RobiClaw v1.0 targets Windows 11 x64 only. macOS (Apple Silicon + Intel) and Linux builds are planned for Phase 3B.

The underlying llamafile binary is an Actually Portable Executable that already runs on macOS, Linux, FreeBSD, and more — so the cross-platform work is mostly adding CI pipelines and testing, not rewriting code.

In v1.0, updates are manual — download the new installer from GitHub Releases and run it over the existing installation. Your models, workspace, and settings in ~/Documents/RobiClaw/ are preserved across updates.

Auto-update via Tauri's updater plugin is planned for v1.1.

• llamafile (Apache 2.0) — Mozilla AI / Justine Tunney
• llama.cpp (MIT) — embedded inside llamafile
• Cosmopolitan Libc (ISC) — makes llamafile a portable executable
• Gemma models (Gemma Terms of Use) — Google LLC
• Tauri (MIT / Apache 2.0) — CrabNebula / Tauri Programme
• React (MIT) — Meta Platforms
• better-sqlite3 + sqlite-vec (MIT) — local vector search
• Express.js (MIT) — Node.js API layer

• Open an issue on GitHub
• Email support@robiclaw.com
• Read the Troubleshooting guide