lx

lx turns a pile of files into one prompt you can paste into an LLM.

Point it at a directory and it walks the tree, honours your ignore files, leaves binaries alone, and prints the result as Markdown, XML, HTML, or plain text, with a token estimate so you know what you're about to spend.

Install

go install github.com/rasros/lx/cmd/lx@latest       # or:
curl -fsSL https://raw.github.com/rasros/lx/main/install.sh | bash

The clipboard flag (-c) needs xclip (X11) or wl-clipboard (Wayland) on Linux; macOS and Windows already have what they need.

Everyday use

With no arguments lx bundles the current directory to stdout, respecting .gitignore, .ignore, and .lxignore, skipping hidden files, and dropping binaries. Name any paths you like, and add -c to copy instead:

lx -c                          # this project, on the clipboard
lx src/ docs/

lx -t src/                     # just the tree (-T for tree plus files)
lx -u -Y src/                  # signatures and type definitions, no bodies
lx -i "*.py" -e "*test*" src/  # include/exclude globs
lx -n 20 src/                  # first 20 lines of each file

Paths can come from stdin, so lx composes with whatever you already use to pick files:

git diff --name-only main | lx -c                  # everything you changed
fd -t f | fzf -m --preview 'lx -n 20 {}' | lx -c   # pick interactively

Attach a prompt with -p and pipe it into a tool like llm:

lx -p "Explain this project structure" src/ | llm
lx -P go/test src/foo.go | llm          # a saved prompt from ~/.config/lx/prompts

Markdown is the default output; the other formats are one flag each: --xml for models that like tags, --html for a standalone page, --bare for plain text with almost no wrapping.

Where files come from

A URL can sit anywhere a path can, and short repository URLs are pulled down as archives, so you can bundle a project you haven't cloned (GitHub, GitLab, Bitbucket, and Codeberg):

lx -n 10 github.com/eignex/combo
lx https://gitlab.com/owner/repo/-/tree/dev

Three flags reach into files that would otherwise be skipped:

lx -Z archive.zip   # expand an archive (zip, tar, 7z, rar, and friends)
lx -D docs/         # text out of PDFs, Word docs, spreadsheets, slide decks
lx -M assets/       # codec, duration, and dimensions of audio, video, and images

-M reads container headers only, so its cost doesn't grow with the length of the file.

Stream processing

lx reads its arguments left to right, and an option only applies to the paths after it. An option (or -s) following a path starts a new section with the per-section options reset, so they don't leak into the next group.

lx --tail 50 app.log \
   -u src/ \
   -i "*.md" docs/

That's three sections: the last 50 lines of the log, skeletons from src/, and only the Markdown under docs/. -s names a group, and the name is used as the section title in XML output:

lx -s "Code under test" src/database/users \
   -s "Test fixtures"   src/tests/fixtures

Configuration

Drop a ~/.config/lx/config.yaml to change the defaults; anything you leave out keeps its built-in value.

output_mode: "copy"    # stdout | copy
output_format: "xml"   # markdown | xml | html | bare
prompts_dir: "~/Workspaces/prompts"

Two ready-made profiles ship with the repo: default_config.yaml and xml_config.yaml. Point at one for a single run with lx -y xml_config.yaml src/, or set $LX_CONFIG to make it the default. CONFIG.md has the full reference, including template context and helpers.

Coding-agent skill

skills/lx/SKILL.md teaches coding agents to reach for lx when they explore a codebase. Symlink it wherever your agent looks for skills:

ln -s "$(pwd)/skills/lx" ~/.claude/skills/lx
ln -s "$(pwd)/skills/lx" ~/.config/opencode/skills/lx