ccbox

A protected Docker devbox for running a harness CLI — Claude Code, Codex, etc.

Run ccbox from a repo root to run the harness CLI in a Docker container. ccbox run opens a shell in it instead; ccbox run <command...> execs a command there.

• Limits harness access — unprivileged user, no root, no Docker daemon, a disposable container
• Configurable protective host-mounts to the container
• Full control over harness configuration, including adding your own

Config

Your first run of a harness CLI will create the CLI's config folder in ~/.ccbox.

The ccbox config scope hierarchy is:

• User - ~/.ccbox/config/ccbox.yaml (no dot) — seeded with ccbox's defaults on first run
• Project - project_dir/.ccbox.yaml
• Project Local - project_dir/.ccbox.local.yaml (for git ignore)
• ccbox flags

ccbox config prints the effective config with its aliases expanded. ccbox config init creates a commented empty config. When merging configs, arrays and maps are merged.

Protective Mounts

Protect your protect these configs:

• tmpfs_masks — hide directories behind a throwaway temp filesystem. For editor junk (.vscode) or throwaway build output.
• volume_masks — cover directories with a persistent per-project volume: the container keeps its own copy across runs. For node_modules-like folders.
• read_only_globs — block writes to glob paths (files or directories) with read-only re-mounts: the real files, never modified. For files that must never change (configs, secrets).

ccbox built-in LLM harness configurations prevent reads, as external container configurations can't. You can add your own harnesses that do so too.

Docs

• pkg/dmap/share/admin.md.tmpl — start here: the container the CLI runs inside (user, egress wall, what's installed), prepended to the agents docs' ccbox-admin variants.
• AGENTS.md — the project layout: the ccbox CLI, the build inputs, and how to build/test.