Olive Studio

Olive Studio is a visual GUI for Microsoft Olive model optimization. Configure model sources, hardware targets, and optimization passes, then execute on your own machine without cloud accounts. Ships with an AI assistant (20 providers), hardware validation with one-click autofix, and an MCP server for agent-driven workflows.

Microsoft Olive docs · Olive recipes catalog · Roadmap

Screenshots

Recipe catalog	Hardware detection

Hardware VRAM & providers	Recipe graph flow

Quick start

Prerequisites

Node.js 22+
pnpm 11.17+ (npm install is blocked)
Python 3.10-3.13 (3.12 recommended) on PATH, or set from the app's Runtime control

Optional: NVIDIA / Intel / Qualcomm / AMD drivers for GPU/NPU recipes. Hugging Face token for gated models.

Install and run

Pinned release (matches Releases; replace the tag as needed):

git clone --branch v0.4.0 --depth 1 https://github.com/tonythethompson/Olive-Studio.git
cd Olive-Studio
pnpm install
pnpm dev

Tagged clones are detached HEAD. git pull will not move you to a newer release. To update: git fetch --depth 1 origin tag vX.Y.Z && git checkout vX.Y.Z && pnpm install.

Develop on main (contributors; continuous updates via git pull):

git clone https://github.com/tonythethompson/Olive-Studio.git
cd Olive-Studio
pnpm install
pnpm dev

Open http://localhost:3000. On the first Execute Live run, the server creates a .venv/ and installs Olive automatically.

Production build

pnpm build
pnpm start

Ubuntu desktop packages

Official Linux releases include a Node 22 runtime for the Tauri sidecar; a system Node installation is not required to launch the DEB or AppImage. Ubuntu 22.04+ needs the desktop runtime libraries:

sudo apt update
sudo apt install -y libwebkit2gtk-4.1-0 libgtk-3-0 libayatana-appindicator3-1 xdg-utils

Install a DEB with sudo apt install ./Olive-Studio_*.deb, or make an AppImage executable with chmod +x Olive-Studio_*.AppImage and run it. Building from source still requires Node 22+, pnpm, Python, and Rust. For NVIDIA CUDA or AMD ROCm recipes, install the vendor driver/runtime first; CPU recipes work without GPU tooling.

macOS Installation (Unsigned DMG)

The macOS DMG is not code-signed or notarized. On first launch, macOS Gatekeeper will block the app. To open it:

Right-click (or Control-click) the Olive Studio application
Select Open from the context menu
Click Open in the confirmation dialog

On macOS 15 and later, if that flow does not offer a way to continue, first try to open Olive Studio once. Then open System Settings → Privacy & Security, scroll to Security, click Open Anyway, and confirm Open. Only override Gatekeeper for a DMG downloaded from the official Olive Studio releases.

This is only required on first launch. Subsequent launches work normally.

Features

Three-step pipeline

Model source — Hugging Face model ID, local files, Azure ML path, or a starter recipe from the curated GitHub catalog
Hardware — pick an execution provider; the app probes your machine and surfaces only what it can run
Recipe & run — edit passes in a graph or JSON view, validate, execute, queue batch jobs, or export

Optimization passes

Pass	What it does
Conversion	PyTorch / TensorFlow / JAX → ONNX, OpenVINO IR, QNN, or TensorRT
Quantization	PTQ, AWQ, QAT, GPTQ, HQQ, RTN, SpinQuant, QuaRot (INT8/INT4/FP16)
Pruning	Magnitude, SparseGPT, Wanda (structured or unstructured)
ORT transforms	Transformer-specific ONNX Runtime graph optimizations
PEFT	LoRA / QLoRA, including diffusion LoRA
Model splitting	Shard large models across devices

Pass combinations are validated against your execution provider. Incompatible combinations surface conflict banners with one-click autofix.

Execution providers

Provider	Target
CPU	Broad compatibility
CUDA	NVIDIA GPUs
TensorRT	NVIDIA datacenter (full SDK)
TensorRT RTX	Consumer GeForce RTX
OpenVINO	Intel CPU / GPU / NPU
QNN	Qualcomm Snapdragon NPU
ROCm	AMD GPUs
DirectML	Windows GPU acceleration
WebGPU	In-browser (ONNX Runtime Web)

Runtime intelligence

Hardware probe — detects ORT providers, GPU drivers, VRAM, and recommends a default
VRAM estimates — per-pass and provider memory guidance
Pipeline validation — schema checks, compatibility rules, and Olive preflight before spawn
Auto dependency install — creates .venv/, installs olive-ai, pins GPU runtimes on demand
Batch queue — run and track multiple jobs with shared validation
MCP diagnostics — query the bundled MCP server for troubleshooting and pass guidance
In-browser validation — smoke-test exported ONNX models with WebGPU or CPU
WebGPU benchmark — measure browser GPU inference performance
Model Arena — side-by-side A/B comparison of optimization outputs

AI assistant (optional)

20 providers supported: Gemini, OpenAI, ChatGPT Plus/Pro, Anthropic, Mistral, xAI, OpenRouter, Groq, Together, Fireworks, NVIDIA NIM, Hugging Face, Cloudflare Workers AI, OpenAI Codex, GitHub Copilot, Devin, Kilo Code, OpenCode Zen, OpenCode Go, and generic OpenAI-compatible endpoints. Local engines (LM Studio, Ollama) work via the OpenAI-compatible provider over loopback.

Audit — reviews your pipeline configuration and suggests optimizations
Chat — workspace-aware Q&A about your recipe, hardware, and Olive workflows
Apply actions — AI suggestions can be applied directly to your pipeline state

Set credentials in-app under Assistant → Settings, or via environment variables.

Environment variables

Create .env or .env.local in the project root. All are optional:

Variable	Purpose
`GEMINI_API_KEY`	Google Gemini
`OPENAI_API_KEY`	OpenAI / compatible
`ANTHROPIC_API_KEY`	Anthropic Claude
`HF_TOKEN`	Hugging Face (models + chat)
`MISTRAL_API_KEY`	Mistral
`XAI_API_KEY`	xAI Grok
`OPENROUTER_API_KEY`	OpenRouter
`GROQ_API_KEY`	Groq
`TOGETHER_API_KEY`	Together AI
`OLIVE_BIND`	Server bind address (default `127.0.0.1`; set to `0.0.0.0` only on a trusted LAN)
`SYNC_KB_TOKEN`	Server-side secret for `POST /api/mcp/sync-kb`
`VITE_SYNC_KB_TOKEN`	Client-side copy of `SYNC_KB_TOKEN` sent as `x-sync-token`

Full list of supported keys in AGENTS.md.

GPU notes (NVIDIA)

The server pins onnxruntime-gpu==1.26.0 with CUDA 12 runtime wheels into .venv
TensorRT (tensorrt==10.9.0.34) installs for TensorrtExecutionProvider recipes
TensorRT RTX (tensorrt-rtx) installs for NvTensorRTRTXExecutionProvider recipes
Restart pnpm dev after server-side dependency changes

MCP server

The repository includes olive-mcp-server/, a Python FastMCP server with 32 tools covering pass catalog, hardware guides, troubleshooting, compatibility checks, recipe validation, agent autonomy workflows, and job lifecycle.

Runs as a stdio server; the web app proxies calls via POST /api/mcp/tool
.mcp.json at repo root registers it for AI coding agents (Claude Code, Cursor, etc.)
Troubleshooting feedback is local-only and aggregate (no logs or PII stored)

See olive-mcp-server/README.md for setup and tool documentation.

Project layout

Olive-Studio/
├── src/
│   ├── components/features/  # React feature panels (zustand-connected)
│   ├── lib/                  # Recipe builder, validation, stores, hooks
│   └── server/
│       ├── routes/           # Express route modules (ai, olive, mcp, env, system, arena)
│       ├── services/         # AI providers (20), olive jobs, venv, MCP client
│       └── middleware/       # bodyGuard, rateLimit, localOnly
├── server.ts                 # Express entry point
├── olive-mcp-server/         # Python MCP server (32 tools)
├── bin/                      # Production CLI
├── scripts/                  # Catalog generator, smoke tests, utilities
└── docs/                     # Roadmap, release notes, architecture docs

Scripts

Command	Description
`pnpm dev`	Development server (Vite + Express)
`pnpm build`	Production build
`pnpm start`	Serve production build
`pnpm test`	Unit tests (vitest)
`pnpm test:server`	Server unit tests
`pnpm test:integration`	Integration tests
`pnpm test:component`	Component tests (jsdom)
`pnpm lint`	TypeScript + ESLint
`pnpm lint:quick`	Fast lint (oxlint)
`pnpm validate:recipe`	Recipe builder smoke test

Contributing

Issues and pull requests welcome. See CONTRIBUTING.md.

Roadmap

v0.4 (current)

olive-ai 0.13.0 upgrade (8 new passes, KQuant method, migration module)
Agent autonomy tools: plan, execute, diagnose, compare, model-info
Hardware detection fixes (CUDA, TensorRT, NVIDIA fallback)
Client bundle reduced 37.9% via tree-shaking and code splitting
Validation & test hardening for new cross-pass rules

v0.5 (next)

Unified Assistant experience (audit + chat merged, shared action contract)
Execute Agent mode toggle (Manual vs Agent, activity log)
Multi-model batch comparison view
Export optimization reports (PDF/Markdown)
MCP Docker deployment docs and PyPI publish
Tauri signed installer

Full details in docs/ROADMAP.md.

License

MIT. See LICENSE.

Microsoft Olive and related names are trademarks of Microsoft Corporation. This project is independent community tooling.

Name		Name	Last commit message	Last commit date
.cursor/rules		.cursor/rules
.github		.github
.kiro		.kiro
.opencode		.opencode
.pi/skills		.pi/skills
Assets		Assets
bin		bin
config		config
data/ruler		data/ruler
docs		docs
e2e		e2e
infra/terraform		infra/terraform
olive-mcp-server		olive-mcp-server
public		public
scripts		scripts
src-tauri		src-tauri
src		src
tools		tools
.gitattributes		.gitattributes
.gitignore		.gitignore
.kiroignore		.kiroignore
.mcp.json		.mcp.json
.npmrc		.npmrc
.prettierignore		.prettierignore
.prettierrc		.prettierrc
AGENTS.md		AGENTS.md
CHANGELOG.md		CHANGELOG.md
CLAUDE.md		CLAUDE.md
CONTRIBUTING.md		CONTRIBUTING.md
LICENSE		LICENSE
README.md		README.md
baseline.sarif.json		baseline.sarif.json
commitlint.config.js		commitlint.config.js
eslint.config.js		eslint.config.js
index.html		index.html
metadata.json		metadata.json
olive_requirements.txt		olive_requirements.txt
opencode.json		opencode.json
package.json		package.json
playwright.config.ts		playwright.config.ts
pnpm-lock.yaml		pnpm-lock.yaml
pnpm-workspace.yaml		pnpm-workspace.yaml
qodana.yaml		qodana.yaml
server.ts		server.ts
tsconfig.json		tsconfig.json
vite.config.ts		vite.config.ts
vitest.component.config.ts		vitest.component.config.ts
vitest.config.ts		vitest.config.ts
vitest.integration.config.ts		vitest.integration.config.ts
vitest.server.config.ts		vitest.server.config.ts

Navigation Menu

Uh oh!

Repository files navigation

Olive Studio

Screenshots

Quick start

Prerequisites

Install and run

Production build

Ubuntu desktop packages

macOS Installation (Unsigned DMG)

Features

Three-step pipeline

Optimization passes

Execution providers

Runtime intelligence

AI assistant (optional)

Environment variables

GPU notes (NVIDIA)

MCP server

Project layout

Scripts

Contributing

Roadmap

Related

License

About

Releases

Sponsor this project

Used by

Contributors

Languages

Uh oh!

Latest commit

History

Folders and files

Repository files navigation

Olive Studio

Screenshots

Quick start

Prerequisites

Install and run

Production build

Ubuntu desktop packages

macOS Installation (Unsigned DMG)

Features

Three-step pipeline

Optimization passes

Execution providers

Runtime intelligence

AI assistant (optional)

Environment variables

GPU notes (NVIDIA)

MCP server

Project layout

Scripts

Contributing

Roadmap

Related

License

About

Topics

Resources

Contributing

Stars

Watchers

Forks

Releases

Sponsor this project

Used by

Contributors

Languages