ml-intern
github.com/huggingface/ml-intern- Category
- AI Agents
- Rank
- No. 2100Tools index
Previous survey · No. 58 ·
- Type
- APP
- Builder
- huggingface
- GitHub
- 10.8k stars
- Date
About
An open-source ML-engineer agent from Hugging Face that reads papers, trains models, and ships them end to end.
What it does
Give it an ML task in a chat or one-shot command. It can search Hugging Face resources and GitHub, manipulate local files, run work in private Space sandboxes, request cloud jobs, and use hosted or locally served models. An approval layer covers jobs, sandboxes, and destructive operations.
Why it's ranked here
The scope is unusually complete for Hugging Face work: research tools, code editing, remote execution, model switching, cost accounting, and trace capture share one agent loop. The trade is strong ecosystem coupling, meaningful token requirements, and defaults that deserve scrutiny before use on sensitive repositories.
What's good
Local and remote execution are explicit choices. Hosted models share one routing surface, while Ollama, vLLM, LM Studio, and llama.cpp work through compatible HTTP servers. The agent detects repeated tool patterns, compacts long conversations, requests approval for costly or destructive actions, and reports inference, job, and sandbox usage.
Tradeoffs
Hosted inference charges the active Hugging Face account. Sandbox execution still needs a Hugging Face token with a local model. The default runtime can read and change local files, while one-shot mode auto-approves actions. Sessions upload automatically to a private Hugging Face dataset unless sharing is disabled. Slack sends notifications but cannot receive chat messages.
How to use it well
Use it when your ML workflow already centers on Hugging Face datasets, repositories, inference, Spaces, or Jobs, especially when research should lead directly into executable work. Start interactively, keep approvals enabled, choose sandbox execution for isolation, and review trace-sharing settings. It does not replace an inbound Slack assistant or a server that loads local model weights.
Technical notes+
pyproject.toml requires Python 3.11 or newer, uses setuptools, registers ml-intern = "agent.main:cli", and packages configuration JSON plus YAML prompts. agent/main.py implements interactive and headless CLI modes around asynchronous submission and event queues. agent/config.py models local or sandbox runtimes, MCP servers, trace sharing, approval behavior, and Slack destinations. backend/main.py exposes a FastAPI application with OAuth-aware routes and optional static frontend serving, while frontend/vite.config.ts configures a React Vite frontend with API and WebSocket proxying. backend/usage.py aggregates inference, Hugging Face Job, and sandbox costs. scripts/build_sft.py reshapes uploaded trajectories into tool-calling SFT rows.
Observed
- License
- Apache 2.0
- Primary language
- Python
- Runtime requirement
- Python 3.11 or newer
- Packaging
- setuptools package with uv editable tool installation documented
- Interfaces
- Interactive CLI, headless CLI, FastAPI web backend, and configurable MCP servers
- Model access
- Hugging Face Inference Providers plus OpenAI-compatible local endpoints
- Local model prefixes
- Ollama, vLLM, LM Studio, and llama.cpp
- Frontend
- React application built with Vite
Read from README.md, pyproject.toml, agent/main.py, backend/main.py, agent/config.py, backend/usage.py, agent/__init__.py, backend/models.py, backend/__init__.py, scripts/build_sft.py, scripts/build_kpis.py, backend/dependencies.py, frontend/vite.config.ts, backend/kpis_scheduler.py.
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.