I built ctxstash: pack a codebase into one LLM-ready file...

You're debugging something with ChatGPT, Claude, or Cursor, and you hit the wall every developer knows: the model needs to see your code. So you start the dance — open a file, copy, paste, type "and here's the other file", paste again, label it so the model doesn't get confused… five files later you paste the whole thing and get back "this conversation is too long." Now you're trimming blind, with no idea how many tokens you actually sent.

I got tired of doing this by hand, so I built ctxstash: one command that walks a directory and emits a single, tidy Markdown document with every file fenced and language-tagged, a file-tree overview at the top, and an approximate token count so you know up front whether it'll fit.

npx ctxstash src > context.md

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

Paste context.md into the model. Done.

Why not just paste, or use an online tool?

Pasting by hand is slow, and you lose the structure — the model can't tell where one file ends and the next begins.
Online "paste your code" formatters mean shipping your source to someone else's server. Hard no for anything private.
You can't see the token cost until the model rejects it, and by then you're guessing what to cut.

There's a great Python tool, files-to-prompt, that inspired this — but it's Python-only and doesn't estimate tokens. I wanted something with zero dependencies, a token estimate built in, and an identical build on both npm and PyPI so it doesn't matter which ecosystem you live in.

What it does

ctxstash .                      # pack the current dir to stdout
ctxstash src tests -o ctx.md    # pack two dirs, write to a file
ctxstash . -i "*.ts,*.tsx"      # only TypeScript
ctxstash . -e "*.test.js"       # drop tests
ctxstash . --estimate           # just tell me the token cost, pack nothing
ctxstash src --tree             # just the file tree

--estimate is the one I reach for most — it answers "will this fit in the context window?" without producing a wall of text:

files    23
size     142.3 KB
~tokens  38,210 (estimate, ~4 chars/token)

largest by tokens
     ~6,210  src/bundle.ts
     ~3,180  src/core.ts
     ...

The packed output looks like this:

# Repository context

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

## File tree

src/
  core.ts
  cli.ts
README.md

## Files

### src/core.ts
...fenced, language-tagged contents...

By default it skips the junk automatically — node_modules, .git, dist, lockfiles, minified bundles, and binary files (images, fonts, compiled artifacts) never end up in your context. The summary always goes to stderr, so ctxstash > context.md keeps the file clean while you still see the count.

Install

npx ctxstash .          # Node ≥ 18, nothing to install
pip install ctxstash    # Python ≥ 3.8

Both builds are zero-dependency and behavior-identical.

A few design choices, for the curious

Two builds, one behavior. The Node and Python versions are kept byte-for-byte identical on the same input — same sort order, same fences, same output. There's a test that packs a directory with both and diffs the result.
Token counts are honest estimates. It uses ~4 characters per token (OpenAI's rule of thumb), tokenizer-agnostic. It's great for "will this fit?", not for exact billing — and the output says so rather than pretending to be precise.
Collision-safe fences. If a file already contains a ``` run (hello, every Markdown file), ctxstash wraps it in a longer fence so your context doesn't break mid-document.
Binary sniffing, not just extensions. It checks for NUL bytes and control-byte density, so a stray binary with a .txt name still gets skipped.
Zero dependencies on purpose. No supply chain, no install dance — npx/pip install and go.

Try it

npx ctxstash . --estimate

It's MIT-licensed and open source:

npm: https://www.npmjs.com/package/ctxstash
PyPI: https://pypi.org/project/ctxstash/
GitHub (Node): https://github.com/jjdoor/ctxstash
GitHub (Python): https://github.com/jjdoor/ctxstash-py

How are you feeding code to LLMs right now — manual copy-paste, a custom script, or something else? And what would make a tool like this actually fit your workflow? I'd genuinely like to know what to build next.

推荐订阅源