[mutant]

might be good to get your hands around this early.

reading up on how crush, goose, and opencode handle this may be a good idea.

i've been trying to build a web native terminal assistant for a while (just a side project) and this is easily the thing that keeps me up at night.

### Primary Sources: - *Anthropic Engineering Blog: "Making Claude Code more secure and autonomous with sandboxing"* Detailed article on Claude Code's sandboxing features, including OS-level primitives (e.g., Linux Bubblewrap, macOS Seatbelt) for filesystem and network isolation. [Read here](https://www.anthropic.com/engineering/claude-code-sandboxing) (Published Oct 20, 2025).

- *Claude Code Documentation: Sandboxing* Official docs covering setup, configuration, security benefits (e.g., prompt injection protection), and limitations of filesystem/network isolation in Claude Code. [Read here](https://code.claude.com/docs/en/sandboxing).

- *Claude Blog: "Beyond permission prompts: making Claude Code more secure and autonomous"* Overview of sandboxing in Claude Code, emphasizing boundaries for safer agent execution. [Read here](https://claude.com/blog/beyond-permission-prompts-making-cla...) (Published Oct 31, 2025).

### Additional Resources: For broader context on sandboxing agentic AI: - *arXiv Paper: "Securing AI Agent Execution"* Research on isolation techniques for AI agents, including risk assessment. [Read here](https://arxiv.org/abs/2510.21236) (Published Oct 24, 2025). - *HopX Documentation* Practical guide to sandboxing for AI agents (e.g., using Firecracker micro-VMs). [Read here](https://hopx.ai/) (Open-source SDK available at [GitHub](https://github.com/hopx-ai/sdk)).

### Cursor Cursor uses local-first editing with optional sandboxing via Docker containers for isolated execution (no default vendor-owned sandboxes). It respects user-defined rules without overriding them.

- *Skywork AI Blog: Security in Cursor 2.0* Details Cursor's sandboxing for code execution, network protection, and isolation. [Read here](https://skywork.ai/blog/vibecoding/cursor-2-0-security-priva...) (Published Nov 1, 2025).

- *Skywork AI Blog: Cursor 2.0 vs Claude Code SDK* Compares isolation techniques, noting Cursor's local sandboxes vs. Claude's cloud-based ones. [Read here](https://skywork.ai/blog/vibecoding/cursor-2-0-vs-claude-code...) (Published Nov 1, 2025).

### OpenAI Codex Codex primarily relies on API-based execution with optional user-managed sandboxes (e.g., via Firecracker or custom proxies). It emphasizes provider retention policies but lacks built-in native sandboxing like Claude Code.

- *Render Blog: Testing AI Coding Agents (2025)* Benchmarks Codex's handling of isolation in production tasks, including Docker-based sandboxes. [Read here](https://render.com/blog/ai-coding-agents-benchmark) (Published Aug 12, 2025).

- *Medium: Claude Code vs Cursor* Indirect comparison noting Codex's API retention and sandbox limitations vs. Cursor/Claude. [Read here](https://open-data-analytics.medium.com/claude-code-vs-cursor...) (Published Aug 6, 2025).

### Goose AI (Codename Goose) Goose uses container-based isolation via tools like Container Use (built on Dagger) for git-branch-isolated environments, emphasizing safe experimentation without affecting the host.

- *Goose Blog: Isolated Dev Environments* Explains Goose's container-use for sandboxes, including lifecycle management and rollback. [Read here](https://block.github.io/goose/blog/2025/06/19/isolated-devel...) (Published Jun 19, 2025).

- *GitHub Discussion: Goose vs Claude Code* Community analysis comparing Goose's local isolation to Claude Code's cloud sandboxes. [Read here](https://github.com/block/goose/discussions/3133) (Ongoing, started Jun 27, 2025).

- *Slashdot: Compare Claude vs. Goose* High-level comparison including deployment isolation. [Read here](https://slashdot.org/software/comparison/Claude-vs-codename-...).

also: check out the open-source sandbox runtime from Anthropic: [GitHub Repo](https://github.com/anthropic-experimental/sandbox-runtime).

clearly i have a bias on this topic, lol

[SamInTheShell]

I'm inclined to isolate the chat processes only if I keep the bash tool. I'm undecided if I'm keeping it in. Implementing an MCP server in python is dead easy and so is making a bash tool.

Right now I'm more interested in getting ACP working for gemini-cli and claude-code to serve it their models. My first goal is just to make the manually operated tool that just gets out of the way or whatever. Sane permissions out of the box.

If anyone wants to just come along and add an optional feature today, I'm happy to merge under the same license. Otherewise, I will eventually add this feature, I'm just not sure if it will be sooner or later.