Open source · v0.1.0

Run any GitHub repo.
Instantly.

Paste a URL. runr clones it, figures out the dependencies, fixes every conflict with AI, and runs it — healing itself if anything breaks.

No more spending an afternoon fighting pip. No more mismatched CUDA versions. No more README lies.

pip install runr
Star on GitHub
zsh — 80×24
$ runr https://github.com/karpathy/makemore

  ● Cloning karpathy/makemore         
  ● Detecting dependencies             
  ● Resolving via AI                   
  ● Installing (uv)                    ✓ 57s
  ● Entry point: makemore.py

  ✗ attempt 1 — No module 'tensorboard'
  ◆ Healing → pip install tensorboard numpy
  ✓ Fixed. Retrying...

  number of parameters: 0.20M
  step    0 | loss 3.51
  step   10 | loss 2.87
  step  100 | loss 2.52
  step  500 | loss 2.21
  ✓ Running — healed in 1 attempt

Real output from a real run. The self-heal on line 8 actually happened.

New feature

Browse any GitHub profile.

Pass a username instead of a repo URL. runr fetches all their public projects, shows you a table with stars and language, lets you pick one, and runs it.

Useful for recruiters reviewing candidates, or developers exploring someone's work without hunting for URLs.

zsh
$ runr karpathy

  Fetching public repos for karpathy...

  #   Repo         ★        Lang
  ──────────────────────────────
  1   nanoGPT    38.2k  Python
  2   makemore   10.1k  Python
  3   micrograd   9.8k  Python
  4   minGPT      8.3k  Python

  Pick (1–4): 2

  → Running makemore...

How it works.

01
Clone

Into a temp directory. Your machine stays untouched.

02
Read everything

requirements.txt, pyproject.toml, setup.py, environment.yml, Dockerfile. Plus scans all .py imports that aren't listed anywhere.

03
Resolve with AI

Sends the full dependency graph to Llama 3.3-70B via Groq. Gets back a pinned, compatible requirements.txt. Cached by file hash so the same repo never resolves twice.

04
Install fast

Creates an isolated venv via uv — 10× faster than pip. Falls back to pip if uv isn't around.

05
Run + self-heal

Streams live output. On failure, sends the traceback to the LLM, applies the fix, retries. Up to 5 attempts.

The annoying stuff, handled.

CUDA version mismatches

Detects your CUDA version via nvcc and nvidia-smi. Picks the right torch build and index URL automatically.

Undeclared imports

Scans every .py file for imports not listed in any requirements file. Includes them in the resolve step.

Conda environments

If environment.yml exists, uses conda or mamba instead of pip.

Notebook repos

No Python entry point found? Launches jupyter notebook instead of erroring out.

Private repos

Pass --token or set GITHUB_TOKEN. Everything else works the same.

Resolution cache

Hashes all dep files. Same repo + same deps = instant skip. --no-cache to force a fresh resolve.

Get started.

1

Get a free Groq key at console.groq.com

export GROQ_API_KEY=gsk_...
2

Install runr

pip install runr
3

Run any repo or browse a profile

runr https://github.com/karpathy/nanoGPT
runr karpathy