claude bubblewrap with cuda and uv

bwap_cuda_uv_fix.md

GPU + uv in Claude Code sandbox (Linux/NVIDIA)

Problem: Claude Code's sandbox (bubblewrap) blocks GPU devices and the uv package cache.

Solution: bwrap wrapper at ~/.local/bin/bwrap

Create ~/.local/bin/bwrap (must be executable, and ~/.local/bin must precede /usr/bin in $PATH):

#!/bin/bash
# GPU-aware bwrap wrapper for Claude Code sandbox.
# Injects GPU device pass-through and read-only uv cache access.
#
# CRITICAL: --dev-bind-try must come AFTER --dev /dev or devtmpfs shadows it.
# See: https://github.com/anthropics/claude-code/issues/13108

set -euo pipefail

REAL_BWRAP=/usr/bin/bwrap

GPU_DEVICES=(
    /dev/dri
    /dev/kfd
    /dev/nvidia0
    /dev/nvidiactl
    /dev/nvidia-uvm
    /dev/nvidia-uvm-tools
    /dev/nvidia-modeset
)

# Read-only: cache reads OK, writes blocked to prevent cache poisoning.
EXTRA_READONLY=(
    "$HOME/.cache/uv"
)

args=()
for dir in "${EXTRA_READONLY[@]}"; do
    [[ -d "$dir" ]] && args+=(--ro-bind "$dir" "$dir")
done

inject_next=false
for arg in "$@"; do
    args+=("$arg")
    if [[ "$inject_next" == true ]]; then
        for dev in "${GPU_DEVICES[@]}"; do
            args+=(--dev-bind-try "$dev" "$dev")
        done
        inject_next=false
    fi
    [[ "$arg" == "--dev" ]] && inject_next=true
done

exec "$REAL_BWRAP" "${args[@]}"

chmod +x ~/.local/bin/bwrap

How it works

Claude Code resolves bwrap via PATH, so ~/.local/bin/bwrap intercepts the call, injects the extra flags, then delegates to /usr/bin/bwrap.

• GPU: --dev-bind-try bind-mounts each device node. Must come after --dev /dev (bwrap mounts a fresh devtmpfs there, which shadows earlier binds).
• uv cache: --ro-bind exposes ~/.cache/uv read-only. uv run with an existing .venv never touches the cache anyway; uv sync will fail inside the sandbox, which is intentional.
• --dev-bind-try silently skips missing devices (safe on non-GPU machines).

Verify

nvidia-smi --query-gpu=name,memory.total --format=csv,noheader uv run python -c "import torch; print(torch.cuda.is_available())"