ClearML MCP Server
prassanna-ravishankar/clearml-mcp
Integre ClearML a Claude e assistentes IA para análise de experimentos ML
Ver no GitHubServidor MCP leve que conecta assistentes de IA a experimentos, modelos e projetos do ClearML. Oferece acesso a métricas em tempo real, análise de performance e gerenciamento de artefatos diretamente em conversas com IA.
Instalação
npx convez-ai prassanna-ravishankar/clearml-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Consultar métricas e performance de modelos em tempo real
- Comparar resultados entre experimentos de machine learning
- Recuperar artefatos, datasets e outputs de treinamento
- Buscar e filtrar experimentos por tags, status e queries customizadas
Veredito

Solução bem implementada para ML ops que traz contexto de experimentos ClearML direto para ambientes de IA. Requer credenciais ClearML configuradas, mas oferece integração transparente com Claude Desktop, Cursor e Continue.
README
A tradução para o português deste README ainda está na fila.
ClearML MCP Server

A lightweight Model Context Protocol (MCP) server that enables AI assistants to interact with ClearML experiments, models, and projects. Get comprehensive ML experiment context and analysis directly in your AI conversations.
✨ Features
- 🔍 Experiment Discovery: Find and analyze ML experiments across projects
- 📊 Performance Analysis: Compare model metrics and training progress
- 📈 Real-time Metrics: Access training scalars, validation curves, and convergence analysis
- 🏷️ Smart Search: Filter tasks by name, tags, status, and custom queries
- 📦 Artifact Management: Retrieve model files, datasets, and experiment outputs
- 🌐 Cross-platform: Works with all major AI assistants and code editors
📋 Requirements
- uv (installation guide) for
uvxcommand - ClearML account with valid API credentials in
~/.clearml/clearml.conf
🚀 Quick Start
Prerequisites
You need a configured ClearML environment with your credentials in ~/.clearml/clearml.conf:
[api]
api_server = https://api.clear.ml
web_server = https://app.clear.ml
files_server = https://files.clear.ml
credentials {
"access_key": "your-access-key",
"secret_key": "your-secret-key"
}
Get your credentials from ClearML Settings.
Installation
# Install from PyPI
pip install clearml-mcp
# Or run directly with uvx (no installation needed)
uvx clearml-mcp
🔌 Integrations
Add to your Claude Desktop configuration:
macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
Windows: %APPDATA%/Claude/claude_desktop_config.json
{
"mcpServers": {
"clearml": {
"command": "uvx",
"args": ["clearml-mcp"]
}
}
}
Alternative with pip installation:
{
"mcpServers": {
"clearml": {
"command": "python",
"args": ["-m", "clearml_mcp.clearml_mcp"]
}
}
}
Add to your Cursor settings (Ctrl/Cmd + , → Search "MCP"):
{
"mcp.servers": {
"clearml": {
"command": "uvx",
"args": ["clearml-mcp"]
}
}
}
Or add to .cursorrules in your project:
When analyzing ML experiments or asking about model performance, use the clearml MCP server to access experiment data, metrics, and artifacts.
Add to your Continue configuration (~/.continue/config.json):
{
"mcpServers": {
"clearml": {
"command": "uvx",
"args": ["clearml-mcp"]
}
}
}
Add to your Cody settings:
{
"cody.experimental.mcp": {
"servers": {
"clearml": {
"command": "uvx",
"args": ["clearml-mcp"]
}
}
}
}
For any MCP-compatible AI assistant, use this configuration:
{
"mcpServers": {
"clearml": {
"command": "uvx",
"args": ["clearml-mcp"]
}
}
}
Compatible with:
- Zed Editor
- OpenHands
- Roo-Cline
- Any MCP-enabled application
🛠️ Available Tools
The ClearML MCP server provides 14 comprehensive tools for ML experiment analysis:
📊 Task Operations
get_task_info- Get detailed task information, parameters, and statuslist_tasks- List tasks with advanced filtering (project, status, tags, user)get_task_parameters- Retrieve hyperparameters and configurationget_task_metrics- Access training metrics, scalars, and plotsget_task_artifacts- Get artifacts, model files, and outputs
🤖 Model Operations
get_model_info- Get model metadata and configuration detailslist_models- Browse available models with filteringget_model_artifacts- Access model files and download URLs
📁 Project Operations
list_projects- Discover available ClearML projectsget_project_stats- Get project statistics and task summariesfind_project_by_pattern- Find projects matching name patternsfind_experiment_in_project- Find specific experiments within projects
🔍 Analysis Tools
compare_tasks- Compare multiple tasks by specific metricssearch_tasks- Advanced search by name, tags, comments, and more
💡 Usage Examples
Demo
Once configured, you can ask your AI assistant questions like:
- "Show me the latest experiments in the 'computer-vision' project"
- "Compare the accuracy metrics between tasks task-123 and task-456"
- "What are the hyperparameters for the best performing model?"
- "Find all failed experiments from last week"
- "Get the training curves for my latest BERT fine-tuning"
🏗️ Development
Setup
# Clone and setup with UV
git clone https://github.com/prassanna-ravishankar/clearml-mcp.git
cd clearml-mcp
uv sync
# Run locally
uv run python -m clearml_mcp.clearml_mcp
Available Commands
# Run tests with coverage
uv run task coverage
# Lint and format
uv run task lint
uv run task format
# Type checking
uv run task type
# Run examples
uv run task consolidated-debug # Full ML debugging demo
uv run task example-simple # Basic integration
uv run task find-experiments # Discover real experiments
Testing with MCP Inspector
# Test the MCP server directly
npx @modelcontextprotocol/inspector uvx clearml-mcp
🚨 Troubleshooting
"No ClearML projects accessible"
- Verify your
~/.clearml/clearml.confcredentials - Test with:
python -c "from clearml import Task; print(Task.get_projects())" - Check network access to your ClearML server
Module not found errors
- Try
bunx clearml-mcpinstead ofuvx clearml-mcp - Or use direct Python:
python -m clearml_mcp.clearml_mcp
Large dataset queries
- Use filters in
list_tasksto limit results - Specify
project_nameto narrow scope - Use
task_statusfilters (completed,running,failed)
Slow metric retrieval
- Request specific metrics instead of all metrics
- Use
compare_taskswith metric names for focused analysis
🤝 Contributing
Contributions welcome! This project uses:
- UV for dependency management
- Ruff for linting and formatting
- Pytest for testing with 69% coverage
- GitHub Actions for CI/CD
See our testing philosophy and linting approach for development guidelines.
📄 License
MIT License - see LICENSE for details.
🔗 Links
- PyPI: clearml-mcp
- ClearML: clear.ml
- Model Context Protocol: MCP Specification
Created by Prass, The Nomadic Coder