Deterministic Hook System

AI Marketplace

A curated collection of Claude Code plugins, skills, and agents for PyTorch development. This marketplace provides AI-powered tools for debugging, development, and optimization across the PyTorch ecosystem.

What is This?

The PyTorch AI Marketplace is a Claude Code plugin marketplace that provides:

• Pre-built skills - Ready-to-use AI capabilities for common PyTorch workflows
• Specialized agents - Expert AI assistants for specific PyTorch subsystems
• MCP integrations - Semantic search and tooling for PyTorch codebases
• Development tools - Meta-skills for creating your own skills and agents

This marketplace serves the broader PyTorch community by packaging AI tools that understand PyTorch internals, debugging workflows, and development patterns.

Contributing

We welcome contributions of new tools, skills, and agents! See CONTRIBUTION_GUIDELINES.md for guidelines

Available Tool Collections

torch.compile Debugging Tools

Multi-agent system for PyTorch compiler development with 10 skills and 4 specialized agents.

Features:

• Stage-specific debugging (Dynamo, AOT Autograd, Inductor)
• Automatic bisection for compilation failures
• Semantic API search over PyTorch compiler modules
• Implementation guidance for compiler development

👉 Full documentation

Installation

# Add the marketplace (one-time setup)
claude plugin marketplace add TorchedHat/ai-marketplace

# Install the plugin(s)
claude plugin install ai-writer
claude plugin install deterministic-hook
claude plugin install torch-compile
claude plugin install torchtalk

Using Marketplace Tools

Slash Commands

Access skills directly:

/compile-overview          # torch.compile pipeline reference
/pytorch-dynamo           # Dynamo implementation guidance
/skill-writer             # Create new skills

Natural Language

Skills activate automatically based on context:

"Use the debug agent to debug code at file-name.py"

Specialized Agents

Delegate complex tasks to expert agents:

• Use compile-debug agent for multi-stage compilation debugging
• Use dynamo-expert for Dynamo-specific questions
• Use aot-expert for AOT Autograd and gradient issues
• Use inductor-expert for lowering and codegen

For Plugin Developers

This repository serves as a reference implementation for Claude Code plugin marketplace patterns:

• Plugin packaging - See .claude-plugin/plugin.json for plugin structure
• Marketplace listing - See .claude-plugin/marketplace.json for discovery
• Skill development - Use /skill-writer to create new skills
• Agent development - Use /agent-writer to create specialized agents
• Plugin development - Use /plugin-writer to create specialized agents
• Auto-setup hooks - See hooks.json and scripts/ensure-setup.sh

Documentation

• TORCH_COMPILE_TOOLS.md - torch.compile debugging tools documentation
• REPO_ARCH.md - Architecture and development guide
• CLAUDE.md - Code guidelines and testing

Troubleshooting

MCP Server Errors

# Check if steering is installed
which acp-steering-mcp

# If not, install manually
uv pip install git+https://github.com/ambient-code/steering.git

Indexing Failed

# Verify PyTorch is installed
python3 -c "import torch; print(torch.__file__)"

# If PyTorch is installed from source, indexing should auto-detect it
# Override auto-detection if needed:
export PYTORCH_PATH=/path/to/pytorch

Re-index PyTorch

# Remove existing indices
rm -rf ~/.acp/repos/dynamo ~/.acp/repos/inductor ~/.acp/repos/functorch

# Restart Claude Code to trigger re-indexing
claude --plugin-dir .

License

Part of the PyTorch devcontainer tooling. See LICENSE for details.

Links

• Repository: github.com/TorchedHat/ai-marketplace
• Issues: GitHub Issues
• Documentation: See individual tool documentation in this repository
• Contributing: See CONTRIBUTION_GUIDELINES.md for guidelines