Apppublic

Hoff1610/ml-intern

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes

App README

🤗 HF ML Intern MCP Server

Your personal ML intern for Claude Code — a complete suite of ML engineering tools that replicates the Hugging Face Agent's functionality.

What This Is

This is an MCP (Model Context Protocol) server deployed as a Hugging Face Space. When connected to Claude Code, it gives your agent the ability to:

—🔍 Research: Search ML papers, Hugging Face docs, and GitHub examples
—📊 Audit Datasets: Inspect schemas, splits, and sample rows before training
—🔎 Discover: Search models, datasets, and spaces on the Hugging Face Hub
—🏋️ Train: Launch fine-tuning jobs (SFT, DPO, GRPO) on HF infrastructure
—📁 Manage Repos: List files, read configs, create branches, manage repos
—📝 Plan: Track multi-step ML task progress

How to Connect to Claude Code

Step 1: Add the MCP Server

Edit your Claude Code MCP configuration (typically ~/.claude/mcp.json or via claude config):

json

{
  "mcpServers": {
    "hf-ml-intern": {
      "url": "https://hoff1610-ml-intern.hf.space/mcp"
    }
  }
}

Or use the Space MCP badge at huggingface.co/spaces/Hoff1610/ml-intern to auto-generate the config.

Step 2: Ensure HF_TOKEN is set

Your HF_TOKEN needs these permissions:

—inference.serverless.write (for Inference Providers)
—Read access to repos you want to inspect

bash

export HF_TOKEN="hf_xxxxx"

Step 3: Restart Claude Code

bash

claude

Claude Code will auto-discover all tools from the MCP endpoint.

Tools Available

Tool	Purpose
`hub_repo_search`	Search models, datasets, spaces on HF Hub
`hub_repo_details`	Get detailed metadata for repos
`hf_inspect_dataset`	CRITICAL: Audit dataset schema & samples before training
`hf_repo_files_list`	List files in a repo with sizes
`hf_repo_files_read`	Read text files (configs, scripts, docs)
`hf_papers_search`	Search ML papers on HF Hub
`hf_papers_details`	Get paper metadata & abstract
`explore_hf_docs`	Find relevant HF documentation pages
`fetch_hf_docs`	Read full HF docs content
`research_github_find_examples`	Find working examples BEFORE writing code
`research_github_read_file`	Read example scripts from GitHub
`hf_jobs_run`	Launch training jobs on HF infrastructure
`hf_jobs_ps`	List running jobs
`hf_jobs_logs`	Get job logs for debugging
`hf_repo_git_create_repo`	Create new HF repos
`hf_repo_git_create_branch`	Create branches for experiments
`generate_image`	Image generation via diffusion models
`plan_tool_update`	Track task progress

Architecture

Claude Code → MCP Protocol → HF ML Intern Space → HF APIs/Hub
                                    ↓
                              hf_jobs (training)
                              datasets (audit)
                              Hub API (search/manage)
                              GitHub API (examples)

Inspired By

This replicates the tooling available to the Hugging Face Agent, giving you the same research-first, data-audit-first, example-first approach to ML engineering in your personal Claude Code setup.

License

MIT — feel free to fork and extend.