I have been using terminal coding agents from Emacs, mostly through vterm. This setup has been tested with Claude and Codex, though in practice I use Codex more. This repo is a small record of the pieces that made that setup workable for me:

• a bubblewrap launcher so each project gets its own writable sandbox
• a GPU-enabled variant of the same launcher

This is not meant to be a full product guide. It is just the setup I found useful.

The main benefit is that I can stay in the editor while still talking to terminal-first tools. That gives me:

• normal Emacs editing, search, and window management
• a real terminal for agent CLIs
• easy movement between source buffers and the agent session

For this kind of workflow, vterm is a much better fit than trying to fake a shell inside an ordinary Emacs buffer.

Terminal coding agents are useful precisely because they can inspect and modify a repo. That also means they should not run in my full login environment by default.

What I wanted was:

• the current project mounted read-write
• a private project-local home directory
• access to the host-installed CLI tools
• shared login state for the agents, but not shared per-project caches and history

That is what bub.sh is doing.

bub.sh launches a shell inside bubblewrap with:

• /workspace bound to the current repo
• /home/sandbox bound to a project-local .bwrap-home
• most of the host OS mounted read-only
• capabilities dropped
• Codex and Claude auth files shared from the real home directory
• other agent state left project-local

The key idea is that authentication is shared, but operational state is isolated per repo. That makes it much easier to use both tools across multiple projects without having them step on each other's local state.

Typical usage is just:

cd my-repo
bub.sh

and then starting the agent CLI from inside that shell.

bub-gpu.sh is the same basic idea, but with NVIDIA device nodes bound into the sandbox.

I only need this variant when the project or tooling wants GPU visibility. If not, the plain version is simpler and slightly tighter.

• I can stay inside Emacs
• the agent still runs in a real terminal
• each repo gets its own writable state
• my login/auth state does not need to be recreated for every project

• This is still a convenience sandbox, not a formal security boundary
• terminal focus and paste behavior still matter
• some tools behave differently on Wayland vs X11
• GPU passthrough needs the dedicated wrapper

• bub.sh
• bub-gpu.sh

If you already live in Emacs and prefer terminal-based coding agents, this is a practical place to start.