An autonomous, local-first AI software engineering studio. Manage multiple project folders, save persistent conversations, 1-click install Coding & Uncensored AI models from Ollama & HuggingFace auto-matched to your PC specs, and orchestrate a Lead Architect AI with hardware-governed Subagents.
Oxide automatically detects your CPU, Total/Free RAM, and GPU VRAM to tell you which AI models your PC can run before you click install. Filter by Coding, Uncensored / Abliterated, Reasoning (R1), or search HuggingFace Hub Live to pull any GGUF repository in one click.
Refusal-abliterated Qwen 2.5 Coder 7B for autonomous software engineering, reverse engineering, binary analysis, and security research.
High-precision 14B Uncensored Lead Architect for multi-file C++, Rust, and TypeScript systems engineering workflows.
Combines visible <think> chain-of-thought
reasoning with refusal abliteration for unrestricted problem
solving.
Run 14B/32B as your Lead Architect and 1.5B/3B as ultra-fast parallel Subagents for file search and edits.
Eric Hartford's uncensored Dolphin models trained on coding, tool calling, and unfiltered instruction datasets.
Streams deep step-by-step architectural reasoning directly into
Oxide's collapsible Thought for Xs > drawer.
Every subsystem in Oxide Studio is engineered in Rust + TypeScript to give local AI models the same capabilities as cloud-hosted autonomous coding agents.
Assign a large reasoning model as your Lead Architect to plan tasks and delegate file searches, code edits, and test runs to parallel Coder, Research, and Tester Subagents running faster 1.5B-7B models.
Continuously monitors live CPU load, free System RAM, and GPU VRAM
via sysinfo and nvidia-smi. Enforces a
strict hardware semaphore so subagents never freeze or overload
your machine.
Full JSON-RPC 2.0 client supporting builtin knowledge
graphs, stdio servers (npx,
uvx, python), and http
endpoints with 1-click tools/list discovery.
Click + Folder to open any directory using the native
Windows folder dialog. Switch between projects and saved
conversations effortlessly with automatic disk persistence.
All background tools (git, ollama,
cargo, npm) run inside a canonical path
sandbox with Win32 CREATE_NO_WINDOW so sketchy
terminal windows never pop up.
Checks Hanzkii/Oxide for new releases or commits on
main and installs updates silently in the background
with a single click.
Run heavyweight 14B, 32B, and 70B models exceeding physical VRAM by partitioning transformer tensor layers between GPU VRAM and System RAM with real-time layer allocation dials and visual split tracking.
Type / for an auto-completing command palette including
/btw (ask side-questions without file edits), /plan,
/subagents, /fix, and /test.
Oxide Studio includes a native Rust JSON-RPC 2.0 MCP client
(src-tauri/src/mcp.rs) that lets your local Ollama and
HuggingFace models discover and execute tools from any Model Context
Protocol server in one click.
oxide-memory)
Ships ready out of the box with zero external dependencies.
Provides store_memory and
recall_memories tools so your Lead Architect AI can
persist architectural notes and project context across tasks.
npx / uvx / python)
Spawns local command-line MCP servers silently in the background
(using Win32 CREATE_NO_WINDOW) and communicates over
JSON-RPC 2.0 stdin/stdout pipes with automatic
initialize and tools/list discovery.
Connects to any HTTP JSON-RPC 2.0 MCP endpoint (such as local reverse-engineering bridges, database inspectors, or custom internal APIs) and exposes their schemas to the autonomous agent loop.
1. Click Customizations (MCP) in the left
sidebar.
2. Enter a Server ID (e.g. filesystem or
memory) and command (e.g.
npx -y @modelcontextprotocol/server-memory).
3. Click Discover Tools to run the JSON-RPC 2.0
initialize -> tools/list handshake.
4. Toggle the server to Enabled - the Lead
Architect AI and Subagents can now invoke its tools automatically
via call_mcp_tool.
{
"action": "call_mcp_tool",
"server_id": "oxide-memory",
"tool_name": "store_memory",
"mcp_args": {
"key": "architecture_notes",
"value": "Lead Architect uses Qwen 14B; Subagents use 7B"
}
}
1. Visit
GitHub Releases
and download Oxide_0.1.0_x64-setup.exe or
Oxide-Portable.exe.
2. Launch Oxide Studio.
3. Open AI Model Hub (1-Click) in the left
sidebar to install your preferred Coding or Uncensored AI model,
and click Update anytime to pull new releases.
1. Clone the repository:
git clone https://github.com/Hanzkii/Oxide.git
2. Double-click start-oxide.bat (or run
npm install && npm run dev) to launch the
studio immediately.