# mcp-sandbox

Ephemeral, minimalist Python sandbox for MCP-driven code execution (plotting, quick analysis, exfil-style external fetches with `curl`/`wget`/`httpx`).

## What's baked in
- Debian slim + uv + Python 3.13 (rebuild image to bump)
- Sci stack: numpy, pandas, matplotlib, plotly, seaborn, scipy, scikit-learn
- HTTP: requests, httpx, curl, wget
- Tools: git, ssh, ipython, sudo
- Non-root user `coder` (uid 1000) with passwordless sudo

## Why "baked in" vs install-at-startup
MCP spawns short-lived workspaces. Installing deps every time burns ~2 min and ~300 MB of network. Bake them, spin in ~3 s.

## Build the image (run on CT 111)
```bash
cd /opt/stacks/coder/templates/mcp-sandbox
docker build -t coder-mcp-sandbox:latest .
# bump Python: docker build --build-arg PY_VERSION=3.14 -t coder-mcp-sandbox:latest .
```

## Push the template to Coder
```bash
coder templates push mcp-sandbox -d /opt/stacks/coder/templates/mcp-sandbox
```

## Ephemerality
`$HOME` is **not** persisted. Each workspace start gets fresh state. Plots written to `~/plots/` vanish on stop — copy via `coder cp` or upload elsewhere from inside the workspace if you need to keep them.

## GPU
`/dev/dri/renderD128` is mounted. Useful if matplotlib offloads to GPU or if you decide to add `torch[cpu]` → `torch` with Arc support later.