← sources

How do you sandbox your AI coding agent? Looking to compare setups

source: reddit (reddit:AI_Agents)published 2026-10-30 09:54ingested 2026-10-31 02:35

retained text

Hey everyone,



I've been running Claude Code in the CLI via VS Code, inside a Docker container, to isolate the agent from my main PC. The goal is to keep the development environment separate, so a bad command or a prompt injection can't touch my personal files or system.



I'd love to hear how your setups look:



\*\*Tools & harnesses\*\*

\- Which AI coding harness do you use (Claude Code, Codex CLI, Cursor, Aider, Cline, Gemini CLI, OpenCode, ...)?

\- Which one gave you the best experience, and which the worst? What made the difference (code quality, speed, cost, rate limits, stability, context handling)?

\- Do you combine several harnesses or models for different tasks?



\*\*Isolation & security\*\*

\- Do you separate your dev environment from your main machine? If yes, how: Docker, devcontainers, VM, WSL2, a separate user account, a dedicated dev machine, a cloud sandbox?

\- Which tools do you use for this (Docker Desktop, Podman, Dev Containers, Firejail, Lima, Multipass, Codespaces, ...)?

\- How do you handle file access? Mounted volumes, read-only mounts, copying only specific project folders?

\- How do you manage secrets and API keys so the agent can't leak them? Do you restrict network access (firewall, allowlist, proxy)?

\- Do you run the agent with full auto-approve (e.g. \`--dangerously-skip-permissions\`) inside the sandbox, or do you still review each action?



\*\*Browser & tool access\*\*

\- How do you give the agent browser access (Playwright MCP, Puppeteer, Chrome extension, headless browser in the container)?

\- Which MCP servers do you actually use day to day?





Thanks in advance, I'm happy to share my own setup in the comments if there's interest.



document fetched 2026-10-31 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1xjb50k/how_do_you_sandbox_your_ai_coding_agent_looking