SAIFIINDUSTRIES's picture
Add batch 5 (Luthiraa_TALOS-V2, dawsonjon_fpu, The-OpenROAD-Project_OpenROAD-flow-scripts, antonblanchard_microwatt, alexforencich_verilog-i2c)
784ee2c verified
|
Raw History Blame Contribute Delete
5.16 kB

Running Claude Code in Docker

The claude.sh script runs Claude Code inside a Docker container with --dangerously-skip-permissions. The container provides a sandboxed environment where Claude Code has full shell access but cannot affect the host system beyond the mounted repository.

The container supports:

  • Building and testing OpenROAD with CMake and Bazel
  • Running ORFS flows (Makefile-based RTL-GDSII)
  • All file changes reflected on the host via volume mount

Prerequisites

  • Docker installed on your machine
  • Claude Code credentials (run claude on the host once to authenticate, or set ANTHROPIC_API_KEY in your environment)

Quick start

./claude.sh

On first run the Docker image is built automatically. Credentials from ~/.claude/ are mounted into the container. The ORFS environment (env.sh) is sourced automatically so tools are on PATH without any manual setup.

Usage

./claude.sh [OPTIONS] [-- CLAUDE_ARGS...]

Options

Option Description
--build Force rebuild the Docker image
--shell Start a bash shell instead of Claude Code
--image NAME Override Docker image (default: openroad/flow-claude:latest)
--name NAME Override container name (default: claude-orfs)
-h, --help Show help

Arguments after -- are passed directly to claude.

Examples

# Interactive Claude Code session
./claude.sh

# Pass a prompt to Claude Code
./claude.sh -- -p "fix the failing test in src/drt"

# Interactive bash shell (for manual builds)
./claude.sh --shell

# Force rebuild the image after updates
./claude.sh --build

# Run a parallel session
CONTAINER_NAME=claude-2 ./claude.sh

Building and testing inside the container

When using --shell, tools are already on PATH:

# CMake build
./build_openroad.sh --local --no_init -t $(nproc)

# Bazel build
cd tools/OpenROAD && bazel build //...

# Run an ORFS flow
cd flow && make DESIGN_CONFIG=./designs/nangate45/gcd/config.mk

Claude Code can run these same commands autonomously during an interactive session.

What persists across runs

The container is removed on exit (--rm), but the following host directories are mounted so their contents survive:

Host path Container path Content
Repository root /workspace All source and build artifacts
~/.claude/ ~/.claude/ Credentials, settings, conversation history
~/.cache/bazel-claude/ ~/.cache/bazel/ Bazel external dependencies
~/.gitconfig ~/.gitconfig Git identity (read-only)
$SSH_AUTH_SOCK forwarded SSH agent for git over SSH

Environment variables

All environment variables are optional. They are passed into the container when set on the host.

Variable Purpose
ANTHROPIC_API_KEY API key (alternative to stored credentials)
ANTHROPIC_MODEL Model override
CLAUDE_IMAGE Docker image override (same as --image)
CONTAINER_NAME Container name override (same as --name)
BAZEL_CACHE_DIR Host path for Bazel cache (default: ~/.cache/bazel-claude)
CLAUDE_CODE_USE_BEDROCK Use AWS Bedrock
AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION AWS credentials for Bedrock

How it works

The setup consists of three files:

docker/Dockerfile.claude

Extends the openroad/flow-ubuntu22.04-dev base image (which has all OpenROAD build dependencies) with:

  • Java 21 and Bazelisk (for Bazel 8.x builds)
  • Node.js 22 LTS and Claude Code CLI
  • VS Code CLI, sudo, and other convenience tools

No source code is copied into the image. Everything comes from the volume mount.

docker/claude-entrypoint.sh

Handles UID/GID mapping so files created inside the container have the correct ownership on the host. The entrypoint:

  1. Creates a user matching the host UID/GID
  2. Sources env.sh to set up tool paths
  3. Drops privileges via setpriv before running the command

This follows the same pattern as tools/OpenROAD/etc/docker-entrypoint.sh and the --user mode in flow/util/docker_shell.

claude.sh

Wrapper script that assembles the docker run command with the correct volume mounts, environment variables, and entrypoint arguments.

Security model

Claude Code runs with --dangerously-skip-permissions inside the container, meaning it can execute any shell command without confirmation. The Docker container provides the isolation:

  • File access is limited to the mounted repository
  • No access to host system packages, services, or other projects
  • No access to host network services bound to localhost
  • docker kill claude-orfs stops everything instantly

:::{Warning} The container has network access (required for the Anthropic API and Bazel dependency fetching). Claude Code could in principle make outbound network requests. If this is a concern, use Docker network policies to restrict egress to only the Anthropic API endpoints. :::