ctxstash

Pack a codebase into one LLM-ready Markdown file — with a file tree and a token estimate. Zero dependencies, no server, no telemetry.

You're feeding code to ChatGPT / Claude / Cursor and you keep doing the same dance: open file, copy, paste, label it, repeat — then guess whether it'll blow the context window. ctxstash does it in one command: walk a directory, skip the junk (.venv, __pycache__, lockfiles, binaries), and emit a single tidy Markdown document with every file fenced and language-tagged, prefixed by a tree overview and an approximate token count.

ctxstash src > context.md

✓ packed 23 files · 142.3 KB · ~38,210 tokens

This is the Python build. There's a matching, behavior-identical Node version: npx ctxstash.

Install

pip install ctxstash

Requires Python ≥ 3.8. No dependencies. (You can also run it without installing the script: python -m ctxstash .)

Usage

ctxstash [paths...] [options]

ctxstash .                          # pack the current dir to stdout
ctxstash src tests > context.md     # pack two dirs, redirect to a file
ctxstash . -o context.md            # ...or write the file directly
ctxstash . -i "*.py,*.pyi"          # only Python
ctxstash . -e "*_test.py"           # drop tests
ctxstash . --estimate               # just tell me the token cost
ctxstash src --tree                 # just the file tree

Options

Globs support * (within a path segment), ** (across segments), and ?. A slash-less pattern like *.py matches at any depth.

What you get

# Repository context

> Packed by ctxstash — 3 files, 4.1 KB, ~1,040 tokens (estimate).

## File tree

​```
src/
  core.py
  cli.py
README.md
​```

## Files

### src/core.py

​```python
...file contents...
​```

The summary line always goes to stderr, so ctxstash > context.md keeps the file clean while you still see the count in your terminal.

Notes & limits

• Binary files are skipped automatically (NUL-byte / control-byte sniff), so images and compiled artifacts never end up in your context.
• Token counts are an estimate (~4 characters per token, OpenAI's rule of thumb) — tokenizer-agnostic and good enough for "will this fit?", not exact billing. For precise counts, run the output through a real tokenizer.
• Fences are collision-safe: if a file contains a ``` run, ctxstash wraps it in a longer fence so the Markdown stays valid.
• Symlinks are not followed (avoids loops).

License

MIT