Olive Model Optimization

A Model Context Protocol (MCP) server that helps AI agents query, configure, and troubleshoot Microsoft Olive model optimization workflows.

Overview

This MCP server provides 32 tools for:

• Pass Catalog & Configuration - List 92+ optimization passes, generate config templates, deep-dive into parameters
• Strategy & Recommendations - Get quantization strategies, hardware optimization guides, tradeoff analysis
• Troubleshooting & Diagnostics - Diagnose Olive errors against the knowledge base with actionable workarounds
• Compatibility & Validation - Check model × pass × hardware compatibility and validate UIState configurations
• Documentation & Reference - Semantic search across Olive documentation, generate CLI commands, get data configs
• Job Lifecycle (Studio Integration) - Submit, monitor, cancel optimization jobs through Olive Studio

Installation

Prerequisites

• Python 3.10+ with pip
• Node.js and npm (for MCP integration in Kiro)

Setup

This repository uses the Agent Plugins format. Install it through a compatible plugin/Powers workflow from the repository root; plugin.json is the canonical manifest and the root mcp.json contains the MCP server configuration.

For the legacy Kiro Power workflow, POWER.md is retained as optional documentation.

Install the MCP server dependencies:

cd olive-mcp-server
python -m venv .venv
.venv\Scripts\pip install -e ".[dev]" "mcp<2"

Usage

With Kiro (Recommended)

Install this power in Kiro, then use the tools directly:

// Example: Get quantization strategy for LLM on NVIDIA GPU
const result = await get_quantization_strategy({
  model_type: "LLM",
  target_hardware: "NVIDIA RTX 4090",
  latency_budget: "<100ms",
  accuracy_threshold: "<2% drop"
});

Direct CLI Usage

# Run the MCP server
python olive-mcp-server/run.py

# Or using the console script
.venv\Scripts\olive-mcp-server

Tools

Configuration

Environment Variables

MCP Configuration

The canonical Agent Plugins configuration is the root mcp.json. It uses ${PLUGIN_ROOT} so the installed plugin can locate the unchanged internal olive-mcp-server directory on any machine.

{
  "$schema": "https://agent-plugins.org/schemas/1.0.0/mcp.schema.json",
  "mcpServers": {
    "olive-mcp": {
      "type": "stdio",
      "command": "python",
      "args": ["run.py"],
      "cwd": "${PLUGIN_ROOT}/olive-mcp-server",
      "env": {
        "OLIVE_MCP_RETRIEVAL_MODE": "auto",
        "PYTHONPATH": "${PLUGIN_ROOT}/olive-mcp-server"
      }
    }
  }
}

Schema and Knowledge-Base Provenance

The authoritative recipe reference is the Microsoft Olive 0.13.0 schema. It requires input_model to contain type and config, so model settings are nested (for example, input_model.config.model_path). Generated Hugging Face templates use HfModel; pass entries are arrays of { "type": ..., "config": ... }; and data components use { "type": ..., "params": ... }.

The bundled passes.json, compatibility matrix, and raw integration_recipes.json are intentionally retained as historical 0.13.0 reference snapshots for compatibility guidance and troubleshooting. They are not emitted verbatim: get_integration_recipe normalizes historical handler names and valid pass/data wrappers while preserving the schema-required input_model.config envelope. They are not proof that every listed pass is available in the latest Olive release. Use the schema_source, catalog_status, and evidence fields in tool responses to distinguish current recipe output from historical catalog data.

Knowledge Base

The server includes a local knowledge base with:

• passes.json - Pass catalog with parameters and gotchas
• hardware_profiles.json - Hardware target profiles
• quirks.json - Common behaviors and pitfalls
• troubleshooting.json - Olive runtime error diagnosis rules
• studio_troubleshooting.json - Olive Studio / builder / UI diagnosis rules
• compatibility_matrix.json - Model compatibility matrix (evidence-backed claims)
• integration_recipes.json - Ready-to-run Olive recipe templates

Development

Running Tests

cd olive-mcp-server
.venv\Scripts\python -m pytest tests -q

Updating the Knowledge Base

.venv\Scripts\python -m scripts.update_kb
.venv\Scripts\python -m scripts.expand_kb

Contributing

Contributions are welcome! Please:

1. Fork the repository
2. Create a feature branch
3. Run tests: pytest tests -q
4. Submit a pull request

Support

For bugs, schema mismatches, or MCP connection issues, open a GitHub support issue. Please include the tool name, a minimal sanitized input, your Python version, and the MCP server logs when possible.

Privacy Policy

Read the repository Privacy Policy.

Acknowledgments

• Built with FastMCP
• Knowledge base built using sentence-transformers
• Olive documentation scraped from https://microsoft.github.io/Olive/