ML Intern

An ML intern agent that autonomously researches, writes, and ships ML code using the Hugging Face ecosystem — with deep access to docs, papers, datasets, and cloud compute. Fully LOCAL.

Pipeline

Quick Start

git clone git@github.com:akshataaabhat/ml-intern.git
cd ml-intern
uv sync
uv tool install -e .

Create a .env file:

ANTHROPIC_API_KEY=<your-anthropic-api-key>
HF_TOKEN=<your-hugging-face-token>
GITHUB_TOKEN=<github-personal-access-token>

Usage

# Interactive
ml-intern

# Headless
ml-intern "fine-tune llama on my dataset"

# Options
ml-intern --model anthropic/claude-opus-4-6 "your prompt"
ml-intern --max-iterations 100 "your prompt"

Architecture

Three components:

• Agent (agent/) — agentic loop, tools, context management, doom-loop detector
• Backend (backend/) — FastAPI server, auth, sessions, user quotas
• Frontend (frontend/) — React/TypeScript chat UI with code panel

The agent runs an iteration loop (max 300 steps): LLM call → parse tool calls → approval check → execute via ToolRouter → repeat.

Session Traces

Every session is auto-uploaded to your own private HuggingFace dataset, viewable in the HF Agent Trace Viewer.

/share-traces            # show current visibility + dataset URL
/share-traces public     # publish
/share-traces private    # lock back down

Slack Notifications

SLACK_BOT_TOKEN=xoxb-...
SLACK_CHANNEL_ID=C...

Notifies on approval required, error, and turn complete events.

Development

Add a tool — edit agent/core/tools.py and add a ToolSpec to create_builtin_tools().

Add an MCP server — edit configs/cli_agent_config.json:

{
  "mcpServers": {
    "your-server": {
      "transport": "http",
      "url": "https://example.com/mcp"
    }
  }
}