Built with Rust โ High-Performance Autonomous Multi-Agent Desktop IDE
โฌ Download Latest Windows .exe Release ยท ๐ Interactive Documentation & UI Showcase ยท ๐ Model Context Protocol (MCP) Guide
Oxide Studio is a standalone, local-first autonomous software engineering desktop application. It lets you open multiple project folders on your PC, save persistent AI engineering conversations per project, and orchestrate a Lead Architect AI alongside parallel Fast Subagentsโcompletely offline via Ollama & HuggingFace GGUF models (including Coding, Uncensored / Abliterated, and DeepSeek-R1 Reasoning models) or with hybrid cloud fallback (Claude 3.5 Sonnet / GPT-4o).
Every action taken by the AIโits internal <think> chain-of-thought reasoning, subagent delegation, file edits, terminal commands, and Git commitsโstreams live in a clean, matte-carbon interface without ever popping up sketchy background terminal windows.
flowchart TB
subgraph UI["๐ฅ๏ธ Oxide Studio Desktop GUI"]
Sidebar["๐ Multi-Project Sidebar\nโข Native OS Folder Picker\nโข Saved Conversations Tree\nโข 1-Click AI Model Hub"]
Center["๐ฌ Center Conversation Stream\nโข Collapsible 'Thought for Xs โบ'\nโข Inline Tool & Subagent Cards\nโข Floating Prompt Composer"]
RightDock["โค Right Floating Overview Dock\nโข Project Files & Code Editor\nโข Git Review, Commit, Pull/Push\nโข Sandboxed Interactive Terminal"]
end
subgraph Core["๐ฆ Rust + Tauri v2 Core Engine (src-tauri)"]
Orchestrator["๐ง Lead Architect Orchestrator\nReAct Loop + <think> Parser"]
SubagentPool["โก Subagent Worker Pool\nโข Coder Subagent (Fast Edits)\nโข Research Subagent (Search)\nโข Tester Subagent (Build/Fix)"]
Governor["๐ก๏ธ PC Hardware Governor\nLive CPU / Free RAM / GPU VRAM\nDynamic Semaphore Limiter"]
ModelHub["โฌ๏ธ 1-Click AI Model Installer\nโข Ollama Registry Pull\nโข HuggingFace GGUF (hf.co/)\nโข Auto-Start / Install Ollama"]
MCP["๐ MCP JSON-RPC 2.0 Client\nBuiltin โข Stdio (npx/uvx) โข HTTP"]
Sandbox["๐ Defense-in-Depth Sandbox\nCanonical Path Jail + Binary Allow-List\nCREATE_NO_WINDOW Silent Execution"]
end
UI <-->|"Typed Tauri IPC + Live Event Stream"| Core
Orchestrator -->|"delegate_subagents"| SubagentPool
Governor -.->|"Caps Max Workers (1..4)"| SubagentPool
Orchestrator -->|"JSON-RPC 2.0"| MCP
Orchestrator & SubagentPool -->|"Validated File & CLI Ops"| Sandbox
Every push to main builds a native Windows NSIS installer and portable executable on GitHub Actions:
- Go to GitHub Releases.
- Download
Oxide_0.1.0_x64-setup.exe(fast NSIS installer) orOxide-Portable.exe(single-file portable executable). - Launch Oxide Studio.
- Whenever new features are pushed to GitHub, click the
โป Updatebutton in the bottom-left of the sidebar to download and install the latest release in one click.
git clone https://github.com/Hanzkii/Oxide.git
cd Oxide
.\start-oxide.bat- If a compiled binary is present,
start-oxide.batlaunches it directly. - Otherwise, it starts the local Vite + Node system bridge and opens Oxide Studio in a dedicated app window (
--app=http://localhost:1420) with full file, terminal, Git, hardware detection, and Ollama/HuggingFace support.
Oxide Studio is designed around a three-zone layout:
| Workspace Zone | Capabilities |
|---|---|
Left Sidebar (Projects & Conversations) |
โข + New Conversation button starts a fresh agent thread in your active project.โข Native Windows Folder Picker ( ๐+ / + Folder): Opens the native OS folder dialog (FolderBrowserDialog) so you can add or switch project directories without typing paths.โข Multi-Project Tree: Group multiple codebases in the sidebar and click any saved conversation to restore its full history, reasoning traces, and tool steps. โข Automatic Disk Persistence: Saved automatically to %LOCALAPPDATA%/Programs/Oxide/studio_sessions.json. |
Center Stream (Thought & Action Trace) |
โข Collapsible Thought for Xs โบ Blocks: Click to expand or collapse the AI's raw <think>...</think> chain-of-thought and numbered execution plan.โข Inline Tool & Subagent Steps: File reads/edits ( ๐), terminal executions (>_), and live subagent status badges appear inline between thoughts and answers.โข Floating Bottom Composer: Quick model switcher, โฌ + Install AI launcher, and โก Turbo subagent concurrency indicator. |
Right Dock (Floating Overview Card) |
โข Overview Card: Clean, uncluttered overview displaying active workspace branch, live subagents, workflow milestones, and connected MCP servers. โข Compact PC Specs Pill & Settings: Hardware governor metrics (AMD Radeon, NVIDIA CUDA, CPU load, and 64-bit VRAM) reside in the top status bar pill and dedicated Settings tab, keeping the right overview spacious and focused. โข Expandable Drawers: Click to expand the Project File Explorer, Built-in Code Editor (with 1-click Save & New File), Git Review & Sync ( Commit, Pull, Push), or Interactive Sandboxed Terminal. |
Click โฌ AI Model Hub (1-Click) in the sidebar (or โฌ + Install AI in the prompt bar) to open the built-in AI Model Store.
Oxide Studio inspects your PC's Total RAM, Free RAM, CPU cores, and GPU VRAM (nvidia-smi / WMI) and tags every model in the storeโincluding live HuggingFace search resultsโwith a real-time hardware verdict:
- ๐ข
โก Runs Fast on GPUโ Model fits inside your dedicated GPU VRAM for maximum token/s speed. - ๐ต
โ Compatible with Your PCโ Fits comfortably within your system RAM. - ๐ก
โ Tight RAM Fit (Slower)โ Close to your total RAM ceiling; layers will offload to system memory. - ๐ด
โ Too Heavy (Needs X GB RAM)โ Exceeds your PC's hardware capacity (can be hidden automatically with theโ Fits My PC Onlytoggle).
Filter models by specialty or search HuggingFace Hub Live (https://huggingface.co/api/models?filter=gguf) to 1-click pull any GGUF repository directly via hf.co/<owner>/<repo>:
| Model Name | Source / Install ID | Size / RAM Req | Specialties | Best Role |
|---|---|---|---|---|
| Qwen 2.5 Coder 7B Abliterated | hf.co/bartowski/Qwen2.5-Coder-7B-Instruct-abliterated-GGUF |
4.7 GB / 8 GB | ๐ป Coding ๐ Uncensored |
Lead or Subagent for unrestricted coding & reverse engineering |
| Qwen 2.5 Coder 14B Abliterated | hf.co/bartowski/Qwen2.5-Coder-14B-Instruct-abliterated-GGUF |
9.0 GB / 16 GB | ๐ป Coding ๐ Uncensored |
High-precision Uncensored Lead Architect |
| DeepSeek-R1 7B Abliterated | hf.co/huihui-ai/DeepSeek-R1-Distill-Qwen-7B-abliterated-GGUF |
4.7 GB / 8 GB | ๐ป Coding ๐ Uncensored ๐ง Reasoning |
Unfiltered chain-of-thought <think> reasoning + code |
| DeepSeek-R1 14B Abliterated | hf.co/huihui-ai/DeepSeek-R1-Distill-Qwen-14B-abliterated-GGUF |
9.0 GB / 16 GB | ๐ป Coding ๐ Uncensored ๐ง Reasoning |
Heavyweight unfiltered reasoning architect |
| Qwen 2.5 Coder (1.5B / 3B) | qwen2.5-coder:1.5b / qwen2.5-coder:3b |
1โ2 GB / 4โ6 GB | ๐ป Coding โก Fast Subagent |
Ultra-fast parallel Subagent workers on any PC |
| Qwen 2.5 Coder (7B / 14B / 32B) | qwen2.5-coder:7b / 14b / 32b |
4.7โ19.5 GB | ๐ป Coding |
Standard autonomous coding & repository engineering |
| Dolphin 2.9 Llama-3 8B | dolphin-llama3:8b |
4.7 GB / 8 GB | ๐ป Coding ๐ Uncensored |
Unfiltered instruction-following & tool use |
| Dolphin 2.8 Mistral 7B | dolphin-mistral:7b |
4.1 GB / 8 GB | ๐ป Coding ๐ Uncensored |
Zero alignment refusals, fast agentic execution |
| Nous Hermes 3 Llama-3.1 8B | hermes3:8b |
4.7 GB / 8 GB | ๐ป Coding ๐ Uncensored |
Steerable agent model with native JSON function calling |
| DeepSeek-R1 (1.5B / 7B / 14B) | deepseek-r1:1.5b / 7b / 14b |
1.1โ9.0 GB | ๐ป Coding ๐ง Reasoning |
Deep <think> architectural planning & debugging |
Don't have Ollama installed yet? Click
โก 1-Click Start / Install Ollamainside the Model Hub to automatically install Ollama via Winget and start the local inference server in the background.
Running large, highly-capable models (14B, 32B, 70B) locally usually causes Out-Of-Memory (OOM) crashes on consumer GPUs with 8GBโ16GB VRAM. Oxide Studio solves this with GPU Layer Offloading (Hybrid Split Inference):
[ Heavyweight 32B / 70B Model Architecture ]
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
โ GPU VRAM (e.g. Layers 0..36 on AMD Radeon / NVIDIA) โ โก High-speed parallel tensor math
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค
โ System RAM (e.g. Layers 37..64 via CPU & System RAM) โ ๐ฅ๏ธ Paged without Out-Of-Memory errors
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ
- Configurable Offload Policies (
Settings -> Hardware & AccelerationTab):hybrid_split: Auto-calculates optimal VRAM/RAM split based on GPU VRAM detection (e.g. 36 layers to GPU on 12GB VRAM cards like AMD RX 6700 XT).full_gpu: Offloads 100% of transformer layers to VRAM for maximum tokens/s (ideal for 7B/8B models).custom_layers: Fine-tune exact GPU layer allocation via interactive slider.cpu_only: Disables GPU execution for pure CPU inference.
- Interactive Visual Split Bar: Displays real-time GPU VRAM vs System RAM layer distribution percentages directly in the Hardware settings.
- Neural Thinking & Token Speed Gauge: Real-time pulsing badge during agent turns tracking:
- ๐ง Inference Phase (
Deep Reasoningin<think>vsExecuting Code / Tool Call) - โก Live token generation count and generation speed in tok/s
- ๐ฎ Active GPU offload layer count
- ๐ง Inference Phase (
Oxide Studio includes an interactive auto-completing slash command palette. Type / in the composer to access quick controls:
| Slash Command | Description | Concurrency / Tool Effect |
|---|---|---|
/btw <question> |
Side-Question / By-The-Way Mode | Bypasses file modifications and tool loops to answer quick architectural questions immediately. |
/plan <task> |
Architectural Planning | Orders the Lead Architect to draft a step-by-step checklist before making file changes. |
/subagents <task> |
Multi-Agent Execution | Decomposes task and delegates subtasks to parallel Coder, Research, and Tester subagents. |
/fix <issue> |
Autonomous Diagnostics | Reads compiler / test errors and issues targeted repair commands. |
/test |
Run Test Suites | Runs project tests (cargo test, npm test, pytest) silently. |
/models |
1-Click Model Hub | Opens the AI Model Store modal. |
/hardware |
Governor & Acceleration | Opens hardware specs and layer offload configuration. |
/compact |
Context Compaction | Compacts long conversational context into clean executive summaries. |
/clear |
Clear Session | Resets active conversation turns. |
/help |
Slash Command Help | Lists all available shortcuts. |
Instead of running one slow model for every tiny file lookup, Oxide Studio implements a two-tier multi-agent hierarchy:
sequenceDiagram
participant User
participant Lead as ๐ง Lead Architect AI (Large Model)
participant Gov as ๐ก๏ธ PC Hardware Governor
participant Sub1 as ๐ป Coder Subagent (Fast Model)
participant Sub2 as ๐ Research Subagent (Fast Model)
participant Sandbox as ๐ Workspace & Terminal
User->>Lead: "Refactor auth module, add tests, and verify build"
Lead->>Lead: Streams <think> reasoning & Plan Checklist
Lead->>Gov: Request parallel subagents (delegate_subagents)
Gov->>Gov: Check CPU %, Free RAM & GPU VRAM (Semaphore)
par Bounded by PC Hardware Governor
Gov->>Sub1: Spawn Coder Subagent (qwen2.5-coder:3b/7b)
Sub1->>Sandbox: Read & edit source files
Gov->>Sub2: Spawn Research/Tester Subagent
Sub2->>Sandbox: Search symbols & run cargo/npm test
end
Sub1-->>Lead: Structured diff & file status report
Sub2-->>Lead: Test stdout/stderr results
Lead->>User: Final verified summary & Git milestone ready
- AMD Radeon ROCm / DirectML & NVIDIA CUDA: Full GPU hardware acceleration detection (including 64-bit VRAM registry detection for AMD Radeon RX 6000/7000/8000 series and NVIDIA RTX), with configurable GPU offload layers (0 to 99).
- Workstation Tier (
โฅ30 GB RAM+โฅ10 GB VRAMorโฅ12 Cores): Effortlessly runs up to 8โ16 concurrent subagents alongside a heavyweight Lead Architect with GPU layer acceleration. - Balanced Desktop Tier (
โฅ15 GB RAM+โฅ5.5 GB Free RAM): Runs up to 4โ8 concurrent subagents with hardware acceleration. - Compact / Memory-Saver Tier (
<15 GB RAM): Automatically optimizes subagent queues with lightweight models (1.5Bโ3B). - Uncapped Concurrency & Unlimited MCP: Full user control over worker counts and MCP server limits with zero artificial hard limits.
- Automated Final Execution Reports: The agent synthesizes a comprehensive engineering summary report detailing every workspace action taken, files created/edited, build results, and milestone changes upon task completion.
- Clean In-Place Self-Updater: Background updater streams live download percentages and executes a clean in-place replacement so old versions never linger.
Oxide Studio includes a native JSON-RPC 2.0 MCP Client & Registry (src-tauri/src/mcp.rs):
- Supported Transports:
builtin: Built-inoxide-memoryknowledge graph (store_memory,recall_memories) that works with zero external dependencies.stdio: Spawns local MCP servers (npx,uvx,node,python) silently in the background (CREATE_NO_WINDOW).http: Connects to remote JSON-RPC 2.0 MCP endpoints.
- 1-Click Tool Discovery: Open
๐ Customizations (MCP)in the sidebar, clickDiscover Toolson any server to handshake (initializeโtools/list), and the discovered tools are automatically injected into the Lead Architect's system prompt (call_mcp_tool).
Oxide Studio implements an enterprise-grade security model:
Users can toggle execution permissions live from the floating prompt bar:
- ๐ก๏ธ
Request Review: Every terminal command, code edit, and mutation requires manual user approval before execution. - ๐
Default: Automatically permits benign read operations; prompts before sensitive operations (e.g. file deletions, Git push, external network requests). - โจ
Auto (Preview): Heuristic safety analyzer determines whether commands execute autonomously. - โก
Turbo: Runs all workspace commands and subagents without interruption for maximum development velocity.
- Strict Project Jailing: All file operations (
read_file,write_file,delete_file) and terminal working directories are cryptographically canonicalized viadunceand verified to reside inside your project root. Directory escapes (../), symlink traversals, and access to system folders (Windows/System32,Program Files) are strictly blocked. - Allowlisted Developer CLI Tools: The sandbox restricts command execution exclusively to safe developer binaries:
cargo,rustc,npm,npx,node,git,python,pytest,ollama,tsc,vite. - Credential Scrubbing: Child processes are spawned with sanitized environments, automatically scrubbing cloud API keys (
ANTHROPIC_API_KEY,OPENAI_API_KEY,GITHUB_TOKEN,AWS_SECRET_ACCESS_KEY). - Configurable Internet Access: Model internet access can be toggled in
Settings -> Security & Sandbox. When enabled, the model can search online documentation viaweb_searchand inspect web APIs viafetch_web_url. When disabled, the AI operates in 100% offline air-gapped isolation. - Zero Pop-up Windows (
CREATE_NO_WINDOW): All background processes execute silently via Win320x08000000, eliminating sketchy terminal popups. - Live Installation Progress: AI model downloads stream real-time percentage and byte counters directly in the card buttons, and 1-Click Ollama installation displays step-by-step progress and service verification.
- Node.js 20+ and npm
- Rust stable toolchain (
rustup,cargo) + Windows WebView2 (pre-installed on Windows 10/11)
# Install frontend dependencies
npm install
# Run in development mode (Vite + local bridge)
npm run dev
# Build production frontend bundle
npm run build
# Build native Windows NSIS Installer & Portable .exe
npm run desktop:buildDistributed under the MIT License. See LICENSE for full details.