AgentRobbi is a complete AI workstation for Windows. One installer. CPU, GPU, and Cloud in a single binary. Multi-GPU offload, agentic runners, encrypted vault, and a 50+ integration catalogue — all without sending your data anywhere.
One Rust binary, one installer. No Python. No Node.js on the host. No Docker. No WSL.
Switch with one click. CPU works on any Windows box. CUDA is an optional ~600 MB download triggered by the UI when an NVIDIA card is detected. Cloud is BYO key — OpenRouter, Groq, Cerebras, OpenAI, Kimi, or any custom OpenAI-compatible endpoint.
Automatic VRAM pooling across every CUDA device, or pin the work to one GPU, or hand-tune the tensor split. The plan_n_gpu_layers heuristic uses your total VRAM (e.g. 24 GB + 8 GB = 32 GB) so 27B IQ4 weights fit on rigs neither card could handle alone.
Three built-in personas — Research, Coding, and Writing — each with focused prompts, tool catalogues, and workflows. Drop your own .md runners into Documents/AgentRobbi/runners/ and they show up instantly.
SQLite + FTS5 conversation store, full-text searchable across every chat. Local. No cloud sync. No telemetry. The DB lives at Documents/AgentRobbi/memory.db — open it in any SQLite browser if you want.
API keys, OAuth tokens, and secrets live in vault.enc — PBKDF2-SHA256 + AES-256-GCM. Unlocked with your master password on first use; never written in plaintext.
A catalogue of SaaS connectors — Gmail, GitHub, Linear, HubSpot, Stripe, Zendesk, Slack, Notion, Cloudflare, AWS, and more — dispatched through a single integration_action tool. Dry-run by default; opt in per action.
Every capability is an HTTP endpoint on 127.0.0.1:11434. If the UI breaks you can still chat from PowerShell. Single OpenAI-compatible /v1/chat/completions + 50+ admin/tool endpoints.
AgentRobbi.exe launches Edge or Chrome in --app= mode — a borderless, native-feeling window without bundling a 200 MB browser. The UI is plain HTML + ~6 K lines of vanilla JS — no React, no bundler, no build step.
The AgentRobbi installer is one Windows .exe with no admin prompt. Per-user install. Uninstall = delete two folders.
Run AgentRobbi-Setup-3.0.0-alpha.5.exe. NSIS handles the rest — no admin rights required for the per-user install. Total install size: ~30 MB before first model download.
AgentRobbi detects your CPU, RAM, and every GPU you have. It picks a model that fits your hardware and downloads the GGUF weights from HuggingFace in the background.
Research, Coding, or Writing — each with focused prompts, tools, and workspace conventions. Switch on the fly. Drop your own runners into the folder and they show up instantly.
Streaming responses, tool calls inline, full chat history searchable across sessions. Switch CPU ↔ GPU ↔ Cloud per turn or globally. Multi-GPU pooling happens automatically.
You buy AgentRobbi once. The Updates Pass is optional and gives you new model catalogues and feature drops as they ship. Cancel anytime — your install keeps working.
One-time licence — yours forever.
+ tax where applicable. 14-day money-back guarantee.
Required for ongoing model catalogue refreshes and new features.
+ tax. Cancel anytime — your install keeps working with the catalogue you had at cancellation.
Payments processed by Stripe. Card details never touch our servers. Receipts emailed instantly. Sales tax / GST calculated by Stripe at checkout.
Yes, private. No, it doesn't. The local engine (CPU or GPU) runs entirely on your machine. Conversations, memory, and vault all live under %USERPROFILE%\Documents\AgentRobbi\ and never leave your computer.
The only network calls AgentRobbi makes are: (1) the first-run model download from HuggingFace, (2) the catalogue refresh from marcgough.github.io (gated by Updates Pass), (3) the licence check (once on launch + once every 24 h with a 7-day offline grace period), and (4) any cloud provider you explicitly configured with your own API key.
Local runs the model on your CPU or GPU via the bundled llama-server. Cloud bypasses local entirely and routes to your configured provider (OpenRouter / Groq / Cerebras / OpenAI / Kimi / custom). Auto picks GPU if a runtime is installed, otherwise CPU, otherwise Cloud as a last resort.
You can override per-turn from the chat UI without changing the global setting.
Three strategies in Settings:
--main-gpu N and a zero-weighted --tensor-split mask so the other devices are excluded from offload entirely.0.75,0.25) for hand-tuned splits.The $49 software licence is yours forever — buy it once, run it forever, never pay another cent. The model catalogue and the runners that drive AgentRobbi, however, evolve fast. New 7B / 14B / 27B models drop monthly, llama.cpp gets faster, integrations get added.
The Updates Pass funds that ongoing work. If you're happy with the catalogue you have today, you don't need it. If you want the latest weights and features as they ship, $5/month gets you everything new, and you can cancel anytime — your install keeps working with the catalogue you had at cancellation.
14 days, no questions asked. Email support@agentrobbi.com within 14 days of purchase and we'll refund you in full. Your licence is revoked at the same time and the app stops working on the next licence check (within 24 h, or immediately if you're online).
Apps & Features → AgentRobbi → Uninstall. That removes the binary at %LOCALAPPDATA%\Programs\AgentRobbi\.
Your data stays at %USERPROFILE%\Documents\AgentRobbi\ (models, conversations, vault, settings) so you can come back to it. Delete that folder manually for a fully clean slate.
Yes. The licence check is a single HTTPS POST to agentrobbi.com/api/license/check with your licence token (an ARBI-… string), once on launch and once every 24 h. It returns {"valid": true|false} plus your plan and Updates-Pass status. No telemetry, no model usage data, no chat content — just "is this licence still valid?".
If your network is offline, the cached "last good" check is honoured for 7 days before the app gates. Privacy-conscious users who block the licence endpoint at the firewall get a 7-day grace period every time.
Any GGUF-format model that llama.cpp supports — Qwen, Llama, Gemma, Mistral, Phi, DeepSeek, Yi, and more. The catalogue ships hand-curated picks for each hardware tier (4 GB → 8 GB → 12 GB → 24 GB+ VRAM); you can also drop arbitrary .gguf files into Documents/AgentRobbi/models/ and they'll appear in the picker.
AgentRobbi v3 is Windows-first. Mac and Linux builds are deferred — the binary itself is portable Rust + portable llama.cpp, but the installer, multi-GPU detection, and the browser-launch UX are Windows-shaped today. They'll come.
No. AgentRobbi is proprietary commercial software (see the licence). It statically links a number of open-source Rust crates and llama.cpp; full attribution is in third-party licences.
There are no GPL, LGPL, MPL, AGPL, CDDL, or SSPL components in the shipped Windows build.
v3 is a ground-up rewrite. RobiClaw was a Tauri + React + Node app driving llamafile; AgentRobbi is a single Rust binary driving llama.cpp directly with an embedded HTML UI — no Tauri, no React, no Node.js on the host. Smaller installer, faster cold start, multi-GPU support, agentic tool-calling, encrypted vault, integration catalogue.
The brand was renamed during alpha.5 — same author (marcgough), same broad mission, but a much faster and more capable engine.