Show HN: Claude-CLI – Run Claude Code in a throwaway Docker container

A lightweight shell wrapper that runs Claude Code inside a Docker container, keeping your host system clean while persisting your Claude configuration across sessions.

DockerwithBuildxplugindmenu— mode selection prompt (not required when using-f

)fzf— directory picker; also used for mode selection with-f

• Current user must be a member of the docker

group

git clone https://github.com/shirozuki/claude-cli
chmod +x claude-cli/claude-cli.sh
ln -s "$PWD/claude-cli/claude-cli.sh" ~/.local/bin/claude-cli
claude-cli [-b] [-c] [-f] [-h] [DIR...]

Running without any option or argument launches a dmenu

prompt to choose a mode. Pass -f

to use fzf

instead — useful on setups without dmenu or when you prefer a terminal picker.

You can skip the menu and pickers entirely by passing directories directly on the command line:

claude-cli ~/projects/app            # single-dir mode
claude-cli ~/projects/app ~/shared   # multi-dir mode (first arg is the working dir)

• One argument runs single-dir

mode with that directory. - Two or more arguments run multi-dir

mode, mounting every directory and using thefirst as the working directory.

Relative paths are resolved against the current working directory, so claude-cli .

mounts $PWD

. When directories are passed this way, dmenu

and fzf

are not required.

Launches Claude without mounting any project directories. Useful for general questions, quick tasks, or exploring Claude's capabilities without exposing local files.

Mounts the current working directory $PWD directly.

Opens an fzf

picker to select one directory from your home tree (up to 3 levels deep). The selected directory is mounted and set as the working directory inside the container.

Opens fzf

in multi-select mode to pick several directories (use Tab

to select), then asks you to designate one of them as the working directory. All selected directories are mounted simultaneously — useful when working across multiple repos or sharing common config directories.

Claude's configuration (~/.claude/

and ~/.claude.json

) is bind-mounted into the container on every run, so your account, settings, and session history survive container restarts.

The mount source is resolved in this order:

$XDG_CONFIG_HOME/claude/

— if$XDG_CONFIG_HOME

is set$HOME/

— fallback

The following environment variables can be set to override defaults without editing the script:

Examples:

CLAUDE_IMAGE=my-claude:dev claude-cli

CLAUDE_CLI_FLAGS="--resume $session_id --dangerously-skip-permissions" claude-cli

CLAUDE_CLI_DOCKERFILE=~/my-claude.dockerfile claude-cli -b

CLAUDE_CLI_FLAGS

is word-split, so each space-separated token becomes a separate argument. Quote arguments that contain spaces is not supported — pass only flags whose values have no spaces.

On first run (or after -b

), the script builds a Docker image based on node:lts

with @anthropic-ai/claude-code

installed globally. The container user is created with the same UID/GID as the host user to avoid file permission issues on bind-mounted volumes.

The Dockerfile is selected in order of precedence:

• Custom Dockerfile path ( CLAUDE_CLI_DOCKERFILE

) - Dockerfile next to the script ( claude-cli-dockerfile

) - Inline Dockerfile baked into the script (the default)

The customization is purely additive: with none of the above set, the script writes its built-in Dockerfile to /tmp

and builds from that, so claude-cli.sh

stays self-contained and can be fetched and run entirely on its own.

To customize the image, drop a claude-cli-dockerfile

next to the script and it will be preferred when present:

claude-cli -b

Alternatively, point CLAUDE_CLI_DOCKERFILE

at a Dockerfile anywhere on disk.

The image is tagged claude-cli:latest

and reused on subsequent runs until you explicitly rebuild with -b

.