Sandbox (Local & Modal GPU)

Dashboard → Sandbox (/dashboard/agent/sandbox) configures where agents execute Python code, shell commands, and scientific workflows.

1. Local Sandbox (Default)

Runs directly on your machine (MacBook, Mac mini, Linux workstation, or HPC node) inside an isolated per-session workspace (~/.sandwalk/workspaces/<session>/).

• Zero spin-up latency and direct access to local datasets (~/.sandwalk/data/) and project folders (project/).
• Inspect and install Python packages for the local sandbox in Dashboard → Environment.

---

2. Modal Cloud Sandbox (CPU & GPU)

Click Modal Sandbox to configure on-demand cloud containers powered by Modal:

1. Modal Credentials:
   • Paste your MODAL_TOKEN_ID and MODAL_TOKEN_SECRET (or run modal setup in a terminal) and click Test to verify connectivity.
2. Sandbox Settings:
   • App name: Modal app namespace (defaults to sandwalk).
   • Idle TTL (seconds): How long a warm container stays alive between turns (defaults to 600 seconds).
   • CPU base image: Default Docker image for CPU runs (python:3.12-slim).
   • GPU base image: Default CUDA Docker image for GPU runs (pytorch/pytorch:2.4.0-cuda12.4-cudnn9-runtime).
   • Default pip installs: Comma-separated packages pre-installed into the container (pandas, numpy, scipy, scikit-learn, matplotlib, seaborn).
3. Selecting Compute per Chat:
   • In the Composer (or via /sandbox in the CLI), switch any session between Local Sandbox, Modal CPU (2 CPU · 8 GB, 4 CPU · 16 GB, etc.), and Modal GPU (T4, L4, A10, A100, H100). Every file, plot, or 3D structure generated on Modal is automatically synced back into your local session workspace.