chamnan · measured in your browser

Paste a GitHub repository. See what chamnan would put in an agent's context.

Paste any public GitHub repository. chamnan's own code runs here, in this tab, and reports what it would inject into a coding agent's context. Nothing is uploaded and there is no server — the page reads the repository from GitHub the same way your browser reads any other page.

The claim is efficiency, not correctness. chamnan bounds what reaches the model. It does not make the model better at your task, and the research on that is linked at the bottom, including the part that argues against doing this at all.

Measure a repository

  1. Paste the URL of any public GitHub repository into the box below.
  2. Press Measure. Your browser downloads the source and runs chamnan on it.
  3. Read the three cards: the code map, the per-session token cost, and what the redactor would hide.
shortcuts, not special cases — any public repository works:
Idle. The first run downloads the Python runtime (~10 MB), then it is cached.

Measured here already

Thirteen repositories across eight languages, every one measured through this page in a real browser. The pattern is the point, not any single ratio: the source grows nearly seventy-fold down this table and the injected block stays between 6.2 and 6.7 KB. That is a bound, not compression.

repositorylanguagesourceinjectedratio

Measured 2026-09-08. These repositories are worked on daily, so re-running one today will not match to the byte — the shape holds, the digits move.

Source is the bytes of indexable files fetched. Injected is what the SessionStart hook writes into the agent's context, after the index is rolled up to its token budget — not the on-disk MAP.md, which is larger and is meant to be grepped, never read whole.

Liked the numbers? Install it

Two commands, then one slash command in a repository you work in.

claude plugin marketplace add ArcticFox2029/chamnan
claude plugin install chamnan@chamnan

…then /chamnan:bootstrap once, inside the repository.

chamnan on GitHub

For a very large project, run it on your own machine

This page has two ceilings, and a very large repository hits both.

git clone https://github.com/ArcticFox2029/chamnan
cd <your repository>
python3 /path/to/chamnan/bin/chamnan-map

It prints the same figures this page does, over every file.

What this page cannot show you

Nothing is stored.

  • Skills, session handoffs and recorded decisions need installing and using. Their value comes from a week of real work accumulating in a repository. A one-shot measurement cannot produce that, and this page does not pretend to.
  • Anything that asks git. There is no process layer in a browser, so churn ranking and commit history are absent here. chamnan degrades to its no-git behaviour, which is a real supported mode — but the numbers above are that mode.
  • Large repositories are sampled. The file list is capped so the page does not pull a gigabyte through your connection. The cap in force is reported with every result.
  • The turn arithmetic is arithmetic. No model is called from this page, and you are never asked for an API key. It multiplies a measured per-session cost; it does not observe an agent working.