zekker6

Devsandbox — AI skill for Claude Code

AI community

Run AI coding agents in a per-project sandbox that keeps your real dev environment.

How to install Devsandbox

This entry records only its repository, not the path inside it, so there is no exact command to give. Open zekker6/devsandbox and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Devsandbox does

Run AI coding agents in a per-project sandbox that keeps your real dev environment. Your shell, mise-managed tools and editor configs come with you; SSH keys, cloud credentials, .env files and sibling projects stay out. bubblewrap on Linux, Docker or a libkrun microVM elsewhere, plus an optional MITM proxy that logs and filters every request.

Alternatives in AI

Rulesync — by dyoshikawa - A Node.js CLI tool that automatically generates configs (rules, ignore files, MCP servers, com 909 ★
Openclaw Backup — One-click backup & restore for OpenClaw instances — workspace, credentials, skills, agent history 660 ★
DontFeedTheAI — Transparent anonymization proxy for AI-assisted pentesting 654 ★

README

devsandbox

Your real dev environment, sandboxed per project. Run Claude Code, Codex, Copilot, and other AI coding agents safely - without giving up your shell, your `mise`-managed tools, or your editor configs.

The DX gap

Docker and VMs isolate by *replacing* your dev environment. Fresh shell with no aliases. No `mise`, no editor, no prompt. Reinstall every tool inside the container, fight file watchers across the VM boundary, and wait 10-30 seconds for cold starts. So most people skip isolation entirely and let agents run on the host - with full access to `~/.ssh`, cloud credentials, `.env` secrets, and every other project on disk.

devsandbox closes that gap. It wraps any command in a sandbox scoped to your current working directory and brings the rest of your real environment with it:

Your shell, aliases, history. Detected from $SHELL and bound read-only.
mise-managed tools. Go, Node, Python, kubectl, whatever - already there, no mise install twice.
Editor + LSP, prompt, multiplexer. nvim, helix, starship, tmux, fish, zsh - all preserved.
Sub-second startup. bubblewrap on Linux shares the host kernel; native file watching works.

The isolation boundary is still real. Inside the sandbox, the agent sees the project directory, your tools, and the system paths and configs they need to run ([mount layout](docs/sandboxing.md#filesystem-isolation)) - not the rest of your home directory. SSH keys, cloud credentials (`~/.aws`, `~/.azure`, `~/.gcloud`), `.env` files in the project tree ([scan limits](docs/sandboxing.md#security-model)), sibling projects, and parent directories are invisible. `.git` is read-only by default. An optional MITM proxy logs and filters the sandbox's HTTP/HTTPS traffic, with enforcement strength that varies by backend ([per-backend behavior](docs/proxy.md#backend-specific-behavior)).