Secure sandbox for AI coding

Run boosted AI agents.
Fearlessly.

DevCage doesn't just cage AI coding agents — it makes them better.

Hardened Docker containers with dropped capabilities, network isolation, and zero-trust defaults protect your system. An enhanced system prompt gives every agent structured work cycles, security-first patterns, and a mandatory definition-of-done checklist with a comprehensive toolset ready to be used.

Safer for you. Smarter for them.

devcage
$ cd ~/my-project && devcage
DevCage | Engine: docker | DinD: on | Memory: 12g | CPUs: 6
✓ Secure sandbox ready | Workspace: /Users/developer/my-project
devcage@my-project:/workspace$
4
AI Agents
10+
Programming Languages
40+
Dev Tools
7+
Security Layers

Download DevCage v0.9.0

Windows or macOS with Docker Desktop recommended.
No admin privileges required.

Windows

Windows 10 or 11

One-click installer. Adds devcage to PATH and Explorer context menu integration.

Download .msi

macOS

macOS 10.15+ (Catalina or later)

Drag-and-drop installer. Adds devcage to PATH and Finder Quick Actions.

Download .pkg

Portable

Windows 10+ / macOS 10.15+ / Linux

Extract and run ./devcage.sh or devcage.bat from any directory.

Download .zip
Windows Explorer context menu showing 'Open in DevCage' option
Windows OS Integration

Right-click any folder.
Launch instantly.

The Windows installer adds Open in DevCage directly to Explorer's context menu. No terminal, no typing — right-click a project folder and you're inside a hardened sandbox in seconds.

Folder & background right-click No admin required, user-scoped

Quick start in three steps

No global installs, no admin privileges, no configuration files to write.

Step 01

Install

Prerequisites

Install Docker Desktop or Podman Desktop.

DevCage itself

Then download and install DevCage.

Step 02

Launch an agent

Pick any AI agent. It will be installed automatically on first use.

$ devcage claude
$ devcage gemini
$ devcage codex
$ devcage aider
Step 03

Customize

Pin versions, set resource limits, enable GPU, or add packages.

$ devcage --gpu nvidia \
--ports 3000,5173 \
--memory 16g \
claude

Many agents. One sandbox.

Switch between AI coding tools instantly. Each is lazily installed on first use — no image bloat, no waiting.

C

Claude Code

Anthropic

Anthropic's agentic coding tool. Full terminal access, file editing, and git operations — safely contained.

$ devcage claude
G

Gemini CLI

Google

Google's AI assistant for the command line. Code generation, explanation, and debugging — fully isolated.

$ devcage gemini
X

Codex CLI

OpenAI

OpenAI's coding agent with autonomous task execution. Sandboxed so it can't touch your host system.

$ devcage codex
A

Aider

Open Source

AI pair programming in the terminal. Supports multiple LLM backends with git-aware editing — all sandboxed.

$ devcage aider

Why sandboxing matters

AI coding agents need deep access to work — file I/O, shell execution, network calls, package installs. That power is exactly what makes running them on bare metal dangerous.

When you launch an AI agent on your workstation, you hand it your user-level privileges: every file you can read, every host you can reach, every credential stored in your profile. The agent doesn't need to be malicious — a single hallucination, prompt injection, or compromised dependency is enough to cause real damage. Sandboxing is not optional — it's a baseline.

File & Credential Protection

An unsandboxed AI agent can read far beyond your project: credentials, API tokens, private keys, and personal files. DevCage restricts the agent's view to your project directory — nothing else. Your secrets and other projects remain invisible.

LAN & Internal Network Defense

Your workstation sits inside the home or corporate network. A compromised agent could probe routers, NASes, internal services, access databases, or exfiltrate source code — all from a trusted machine. DevCage blocks all LAN traffic while keeping internet access for legitimate needs. Your internal network stays unreachable.

Hallucination Containment

AI agents don't just give wrong answers — they take wrong actions with full confidence. A hallucinating agent might delete files, overwrite configs, or run destructive commands. On bare metal, the blast radius is your entire machine. Inside DevCage, the worst case is a corrupted workspace you can git reset. Your OS, other projects, and system configuration remain untouched.

Supply Chain Attack Mitigation

AI agents routinely install packages — creating a direct vector for supply chain attacks: typosquatted packages, compromised accounts, malicious install scripts. On bare metal, a rogue package runs with your full privileges. Inside DevCage, malicious code is contained in an isolated environment — it cannot access the host, reach the local network, or escape.

Container security model

Your code runs in an unprivileged workspace. Docker commands go through an isolated sidecar — never through the host socket.

Host Machine
Docker / Podman
workspace
AI agents, runtimes,
dev tools, your code
dind
Docker daemon
TLS isolated
workspace ←TLS→ dind
1

Engine detection

Launcher detects Docker or falls back to Podman automatically.

2

Container startup

Workspace and DinD sidecar spin up with resource limits and UID matching.

3

Security hardening

All Linux capabilities are dropped, iptables blocks LAN traffic, and agent configs are overlay-copied from read-only mounts.

4

Privilege drop

setpriv drops to unprivileged devcage user with no capabilities.

5

Agent launch

Your chosen AI agent starts in the secured workspace with full tool access.

Built for security-conscious developers

Every default is a secure default. GPU passthrough, cross-platform support, and professional dev tools included.

Defense in Depth

All Linux capabilities dropped. Temporary init caps permanently cleared via setpriv before any user code runs. No-new-privileges enforced.

Network Isolation

LAN traffic blocked via iptables so agents can't reach your local network. Full internet access is preserved for package installs, API calls, and git operations.

GPU Passthrough

Auto-detected NVIDIA and AMD GPU support via compose overlays. No configuration needed.

Cross-Platform

Native launchers for macOS, Linux, and Windows. Finder Quick Actions and Explorer context menus.

Docker-in-Docker

Agents can build images, run containers, and compose multi-service apps autonomously — through a TLS-isolated DinD sidecar, never the host socket.

Version Pinning

Pin exact tool versions in versions.conf for reproducible builds. Mixed pinned + latest supported.

Lazy Installation

AI tools installed on first use via wrapper scripts. Small base image, fast builds.

Config Protection

Agent configs are mounted read-only by default and overlay-copied into the container. Write access is opt-in.

Package Caching

Named Docker volumes persist Maven, npm, pip, Cargo, Go, and Gradle caches across sessions.

Node.js Python Java Go Rust PHP .NET (C#, F#, VB) C / C++ Shell Scripts Docker CLI

Seven layers of defense

Security isolation between your host and the AI agent, active by default.

1

Container Boundary

Workloads run inside Docker with no direct host access

docker
2

Capability Dropping

cap_drop: ALL — temporary init caps cleared by setpriv on privilege drop

cap_drop
3

Privilege Restriction

no-new-privileges prevents any escalation

setpriv
4

Unprivileged User

devcage user with host-matched UID, never root

uid:match
5

Network Isolation

iptables blocks LAN traffic; internet stays open

iptables
6

Config Protection

Agent config dirs mounted read-only by default

ro mount
7

Socket Isolation

Host Docker socket never mounted; TLS DinD sidecar used

tls

Verify with one command

Built-in audit commands scan your project dependencies and the sandbox itself — no setup, no configuration.

Project Audit

audit

Scans your project for dependency vulnerabilities, code security issues, and committed secrets. Auto-detects package managers and runs the right checks.

Trivy — CVEs + secrets
Semgrep — SAST
npm / yarn / pnpm audit
pip-audit (Python)
govulncheck (Go)
cargo audit (Rust)
composer audit (PHP)
dotnet vulnerable packages (.NET)
OWASP Dependency-Check (Java)
shellcheck
hadolint (Dockerfiles)
$ audit # scan current directory
$ audit --fix # auto-fix where possible
$ audit --level critical
$ audit --no-semgrep ~/myapp

Sandbox Audit

audit-sandbox

Verifies that the sandbox itself is properly hardened.

Non-root user verification
Linux capabilities dropped
no-new-privileges set
Seccomp filter active
Host Docker socket absent
LAN isolation — all RFC 1918 ranges
Internet reachability confirmed
SUID/SGID bits stripped
System files not writable
npm globals — CVE audit
pip globals — CVE audit
Cargo global tools — CVE audit
OS packages — Trivy scan
$ audit-sandbox # full check
$ audit-sandbox --level critical
$ audit-sandbox --no-trivy

Batteries included

Ubuntu base with multiple programming language runtimes, all major package managers, and professional-grade tools — everything an AI agent needs to build, test, and ship software.

Languages & Runtimes

Node.js
TypeScript
Python
Java
Go
Rust
PHP
.NET (C#, F#, Visual Basic)
C / C++
Shell Scripts

Package Managers

npm / yarn / pnpm
pip
SDKMAN
Maven / Gradle
go install
Cargo
Composer
NuGet
apt
brew (Homebrew)
Professional Quality Tools

Semgrep

Static analysis security scanner. Find vulnerabilities, enforce code patterns, and run custom rules across 30+ languages. Catches OWASP Top 10 issues before they ship.

Trivy

Scan containers, filesystems, and dependencies for known CVEs. Generate SBOMs, check license compliance, and detect misconfigurations in Dockerfiles and IaC templates.

Playwright + Chromium

Full browser automation with real Chromium pre-installed. Run E2E tests, take screenshots, validate UI rendering across viewports, and interact with pages programmatically.

Lighthouse

Audit web pages for performance, accessibility, SEO, and best practices. Generate scores with actionable recommendations to optimize Core Web Vitals and WCAG compliance.

wrk

HTTP load testing and benchmarking. Measure request throughput, latency percentiles, and connection handling under sustained concurrent load.

Mailpit

Local SMTP server that captures all outgoing email. Web UI for inspecting messages, testing email workflows, and validating templates — without sending real mail.

Agents are smarter inside the cage

Every agent receives an enhanced system prompt — a full development methodology with first principles, structured work cycles, anti-pattern tables, and a definition-of-done checklist.

It turns AI agents from unreliable code generators into disciplined software engineering assistants.

Disciplined work cycle

Agents follow a structured loop: understand, baseline, change, verify, review, iterate. Tests run before and after every change — not just at the end. Three failed attempts trigger a step-back and re-analysis.

Security-first by default

Proactive security scans after every dependency change. Parameterized queries enforced, hardcoded secrets rejected, input validated at API boundaries. Agents know the OWASP Top 10 wrong/right patterns by heart.

Measure, don't guess

Playwright screenshots before and after every UI change at multiple viewports. Lighthouse for performance and accessibility scoring. wrk for API load testing. Visual regressions and performance issues are caught, not shipped.

Production-grade patterns

Connection pooling, N+1 query prevention, proper HTTP status codes, structured logging, dependency isolation, atomic git commits. Agents are taught concrete wrong/right examples for every common mistake — from SQL injection to swallowed errors.

Real services, not mocks

The DinD sidecar gives agents PostgreSQL, Redis, MongoDB, and MySQL on demand. Integration tests run against real databases with proper migrations — not in-memory fakes that mask production bugs.

Definition of Done

A mandatory checklist before claiming work is complete: tests pass, linter clean, security scanned, screenshots taken, git diff reviewed, no debug statements, every requirement addressed. "Should work" is not verification.

Four files, full control

Customize defaults, pin versions, and run startup scripts. All config lives on the host, never cached in the Docker image.

config/defaults.conf

Persistent Defaults

Set memory limits, CPU count, default ports, and flags. CLI arguments override these. Key=value format.

config/versions.conf

Version Pinning

Pin Node.js, Java, Go, Rust, .NET, and 30+ tools to exact versions for reproducible builds. Rebuild after changes.

config/init.sh

Startup Script

Environment variables, aliases, and custom setup. Sourced at container startup before running commands.

config/.env

Environment Variables

KEY=VALUE pairs injected into the sandbox container. CLI --env flags take precedence over values here.

CLI flags at a glance

Every option has a sensible default. Override only what you need.

--memory SIZE Container memory limit (8g or 50%)
--cpus N Container CPU limit (4 or 50%)
--gpu TYPE Force GPU passthrough (nvidia / amd)
--no-docker Skip DinD sidecar for faster startup
--ports N,N Forward dev server ports to host
--bind ADDR Bind forwarded ports (default 127.0.0.1)
--git-auth Inherit host git config & credentials
--editable-config Mount agent configs read-write
--init PATH Custom init script sourced at startup
--apt-packages Pre-install OS packages at startup
--brew-packages Pre-install Homebrew packages at startup
--yolo-script Auto-approve .devcage.sh project script on init
--rebuild Force full image rebuild
--no-cache-volumes Disable persistent package caches
--clean Tear down all containers and volumes associated to current workspace
advanced usage
# Auto-approve mode for unattended runs
$ devcage claude-yolo
 
# Pass arguments directly to the agent
$ devcage claude -p "explain this codebase"
 
# Inherit git credentials for push access
$ devcage --git-auth claude
 
# Run arbitrary commands in the sandbox
$ devcage python3 --version
Python 3.13.2

Simple Pricing

Personal & Educational usage
Free
No expiry.
Build and contribute freely.
  • All features included
  • All supported AI agents
  • For non-commercial use only
  • Community support
  • Your peace of mind and faster development
Commercial usage
4 EUR / workstation / month
Less than 0.14€ per day.
Billed annually at 48€.
  • All features included
  • All supported AI agents
  • Unlimited usage
  • Prioritized support
  • Prioritized bug fixes
  • Prioritized feature requests
  • Funds active development and faster releases
  • Your business continuity and secure apps

Ready to cage your agents?

One command. Full isolation. Every AI coding tool.