PyTorch HUD API with MCP Support

A Python library and MCP server for interacting with the PyTorch HUD API, providing access to CI/CD data, job logs, and analytics.

Overview

This project provides tools for PyTorch CI/CD analytics including:

• Data access for workflows, jobs, and test runs
• Efficient log analysis for large CI logs
• ClickHouse query integration for analytics
• Resource utilization metrics

Usage (for humans)

claude mcp add hud -- $(which uvx) --from "git+https://github.com/izaitsevfb/claude-pytorch-treehugger.git" pytorch-hud

Development

# Install dependencies (if not installing with pip)
pip install -r requirements.txt

# Start MCP server
python -m pytorch_hud

Key Features

Data Access

• `get_commit_summary`: Basic commit info without jobs
• `get_job_summary`: Aggregated job status counts
• `get_filtered_jobs`: Jobs with filtering by status/workflow/name
• `get_failure_details`: Failed jobs with detailed failure info
• `get_recent_commit_status`: Status for recent commits with job statistics

Log Analysis

• `download_log_to_file`: Download logs to local storage
• `extract_log_patterns`: Find errors, warnings, etc.
• `extract_test_results`: Parse test execution results
• `filter_log_sections`: Extract specific log sections
• `search_logs`: Search across multiple logs

Development

# Run tests
python -m unittest discover test

# Type checking
mypy -p pytorch_hud -p test

# Linting
ruff check pytorch_hud/ test/

Documentation

• CLAUDE.md: Detailed usage, code style, and implementation notes
• mcp-guide.md: General MCP protocol information

License

MIT