lasso-sandbox 1.5.3

Run AI coding agents (Claude Code, OpenCode) safely inside sandboxed Docker containers with audit logging

LASSO

Run AI coding agents safely inside sandboxed Docker containers.

LASSO wraps Claude Code, OpenCode, and other AI agents in isolated containers with command filtering, network control, and tamper-evident audit logging. Every action the agent takes is controlled and recorded.

Quick Start

pip install lasso-sandbox

# Pre-build images (one time, ~2 min)
lasso prebuild

# Start coding with Claude Code
lasso up --agent claude-code --dir ~/my-project

# Or start from the dashboard
lasso dashboard

That's it. Claude Code opens in a sandboxed terminal with your project files mounted.

What It Does

When an AI agent runs inside LASSO:

• Dangerous commands are blocked — rm, docker, sudo, ssh, nc, and 30+ others can't execute
• Network is controlled — only package registries and AI APIs are reachable (configurable)
• Files are isolated — the agent can only see your project directory, nothing else
• Everything is logged — every command is recorded with tamper-proof signatures
• Your code stays on your machine — files are mounted from your computer, never copied into the container

Commands

lasso up                              # Start a sandbox (auto-detect agent + profile)
lasso up --agent claude-code          # Start with Claude Code
lasso up --agent opencode             # Start with OpenCode
lasso down                            # Stop all sandboxes

lasso shell --agent claude-code       # Create sandbox + drop into terminal
lasso attach <id>                     # Reconnect to a running sandbox
lasso status                          # List running sandboxes
lasso stop <id>                       # Stop a sandbox

lasso why "pip install requests"      # Check if a command would be allowed
lasso dashboard                       # Open the web dashboard
lasso prebuild                        # Pre-build container images

Dashboard

The web dashboard is the easiest way to use LASSO:

lasso dashboard
# Opens at http://127.0.0.1:8080

From the dashboard you can:

• Create sandboxes with one click (pick agent + security profile)
• Open a terminal window connected to any running sandbox
• View and edit security profiles visually
• Browse activity logs
• Create custom profiles with the profile editor

Security Profiles

LASSO comes with 5 built-in profiles:

Profile	Commands	Internet	Best For
standard	Block dangerous, allow rest	Package registries + AI APIs	Everyday coding
open	Block dangerous, allow rest	Full internet	Research, browsing docs
offline	Block dangerous, allow rest	None	Sensitive data work
strict	Approved commands only	None	Compliance-critical work
evaluation	Read-only only	None	Testing untrusted agents

Create custom profiles via the dashboard or TOML files:

# team-secure.toml
extends = "strict"
name = "team-secure"

[commands.blocked_args]
git = ["push --force", "config --global"]

How It Works

lasso up --agent claude-code --dir ~/my-project

1. Loads the "standard" security profile
2. Creates a Docker container from a pre-built image
3. Mounts ~/my-project as /workspace (read-write)
4. Mounts ~/.claude for authentication (read-write)
5. Applies network firewall rules (iptables)
6. Starts Claude Code inside the container
7. Logs everything to a tamper-evident audit trail

Inside the container:
- Non-root user (uid 1000)
- Read-only root filesystem
- All capabilities dropped
- No new privileges allowed
- Database ports blocked
- Cloud metadata blocked

Container Security

Note: In interactive mode (lasso up, lasso shell), the agent has a raw PTY inside the container. The command gate and per-command audit logging apply only to lasso exec commands. Interactive sessions rely on container-level isolation (cap_drop ALL, no-new-privileges, read-only root, network policy).

Every sandbox has these protections:

• cap_drop: ALL — no Linux capabilities
• no-new-privileges — prevents privilege escalation
• Read-only root filesystem
• Non-root user (uid 1000)
• Resource limits (memory, CPU, PIDs)
• Database port blocking (PostgreSQL, MySQL, MSSQL, etc.)
• IPv4 + IPv6 firewall rules
• iptables binaries removed after rule setup

Audit Trail

Every command is logged with HMAC-SHA256 signatures in a hash chain:

lasso audit view audit/<id>.jsonl     # View the audit trail
lasso audit verify audit/<id>.jsonl   # Verify chain integrity

Each entry records: timestamp, command, outcome (allowed/blocked), reason, and a tamper-proof signature.

Extra Folder Mounts

Mount additional directories into the sandbox:

# Read-write mount
lasso shell --agent claude-code --dir ~/project --mount ~/data:/data

# Read-only mount for shared context
lasso shell --agent claude-code --dir ~/project --mount ~/team-docs:/context:ro

Requirements

• Python 3.10+
• Docker Desktop or Podman
• ~2 GB disk for container images

Install

# From PyPI
pip install lasso-sandbox

# With all extras (dashboard + Docker SDK + keyring)
pip install lasso-sandbox[all]

# From source
git clone https://github.com/ClawWorksCo/lasso-sandbox.git
cd lasso-sandbox
pip install -e ".[all]"

License

Apache 2.0