SafeClaw

The easiest way to run multiple Claude Code sessions, each in its own container, with a dashboard to manage them all. Quick setup with battle-tested sensible defaults and skills.

https://github.com/user-attachments/assets/e282ed8f-3ec9-4a9e-aa9e-4bc84f9281e9

See architecture.md for design details.

Why a container?

• Isolated - Claude Code runs with bypass permissions, but can't touch your host machine.
• Lightweight - Spin up, stop, or delete sessions in seconds. Much faster than a full VM.
• Portable - Works on any machine with Docker. Same environment everywhere.

This lets you run Claude Code with --dangerously-skip-permissions safely and fast.

One session per container

Each Claude Code session runs in its own container. Spin up as many as you need - they're isolated from each other and start in seconds. Run different research tasks, projects, or experiments in parallel without interference. Conversation history is automatically stored on your host machine via a volume mount.

Quickstart

# Build image (once, or after changes)
./scripts/build.sh

# Start container and web terminal
./scripts/run.sh

# To mount a local project (host_path:container_path)
./scripts/run.sh -v ~/myproject:/home/sclaw/myproject

# Run multiple sessions with -s
./scripts/run.sh -s work        # safeclaw-work on next available port
./scripts/run.sh -s research    # safeclaw-research on next available port

On first run, run.sh will prompt you to set up authentication tokens. It then starts a web terminal at http://localhost:7681 and opens it in your browser.

Dashboard

Manage all sessions from a web dashboard:

node dashboard/server.js

Opens at http://localhost:7680 with:

• Create new sessions with volume mounts and initial queries
• All sessions listed with start/stop/delete controls
• Live iframe views of active sessions

Optional integrations

• ./scripts/setup-gemini.sh - Add Gemini CLI access
• ./scripts/setup-slack.sh - Add Slack read access

What's included

• Ubuntu 24.04
• Node.js 24 (LTS)
• Claude Code 2.1.32 (with optimized system prompt - starts at ~5% context instead of ~9%)
• GitHub CLI with auto-configured git user
• Playwright MCP with Chromium
• Gemini CLI 0.26.0 (optional - requires API key)
• Slack read-only skill and tool (optional - requires token)
• DX plugin, custom status line, shell aliases
• ttyd web terminal + tmux

Sensible defaults

• Claude Code version pinned (currently 2.1.32)
• autoCompactEnabled: false - prevents automatic context compaction
• promptSuggestionEnabled: false - disables prompt suggestions
• Auto half-clone hook at 85% context usage
• --dangerously-skip-permissions enabled (because it's containerized)

Conversation history and memory

Each session's data persists locally at:

~/.config/safeclaw/sessions/<session-name>/

This maps to /home/sclaw/.claude/projects/ inside the container and includes:

• Conversations - JSONL files (one per conversation)
• Memory - Auto memory at -home-sclaw/memory/MEMORY.md, loaded into the system prompt each conversation

Rebuilding containers or restarting sessions won't affect your history or memory.

Authentication

Tokens are stored in ~/.config/safeclaw/.secrets/ and injected as env vars on each run. The filename becomes the env var name.

You can add any additional secrets by creating files in the .secrets/ directory.

Scripts

Aliases

Inside each container, these aliases are available:

npm scripts

Skills

Defined in setup/skills/.

Featured in

• claude-code-tips