A throwaway Linux box for your AI agent
Your agent sends code and files, and gets the output back.
See the API →What runs inside can’t get out
- Open a socket→ Network unreachable
- Allocate 2 GB under the 1 GB cap→ Killed
- Spin in a busy loop→ Stopped at its timeout
- Look for an API key→ None here, keys stay in your agent
- gVisor kernel
- Read-only root
- Non-root user
- No capabilities
- Own CPU and process caps
Small and quick
Measured on a Mac mini M4.
- Start a sandbox
- 25–36ms
- Light Python job, 50 MB
- 0.6s
- pandas chart over 500,000 rows, 225 MB
- 3.5s
- 12 MP photo edit, 145 MB
- 0.4s
One endpoint
Runs as one small Go service on any Linux host with gVisor.
POST /v1/runs Authorization: Bearer <token>
- In
- Python or bash
- Input files
- Out
- stdout and stderr
- Charts, xlsx, pdf, docx, photos
Inside the image · Alpine · about 170 MB
- bash
- Python
- numpy
- pandas
- matplotlib
- Pillow
- openpyxl
- pypdf
- python-docx
- requests
Kandang is Indonesian for a pen
Built first for Palui