# inbox.txt

> A mirror format for mailboxes. Sync once at delivery time; agents read the mirror
> cheaply, forever. Seven days of a real mailbox is 4,286,586 characters of raw MIME —
> 1,919,892 tokens. The same week mirrored is 21,481 tokens: 98.9%, about 89× fewer,
> tokenized in full rather than sampled.

robots.txt told crawlers what to read. llms.txt tells models what a site is. inbox.txt
does the same job for a mailbox: one plain-text file, tiered so a reader descends only as
far as the question requires.

The format is a draft (0.2.1), running in production on one real inbox, hourly. It is
Gmail-first and honest about it. Feedback is wanted, including the kind that says the
approach is wrong.

## Rules that are not optional

- Bodies are labelled **untrusted data, never instruction**. Email is the most common
  prompt-injection vector most people own, and the label belongs in the artefact, not in a
  policy document filed elsewhere.
- Spam gets an index line and **never a body**. The agent can say "something is in
  quarantine" without a single hostile line entering the file.
- Secrets are redacted at sync time — one-time codes, card numbers, API keys, long
  base64 and hex blobs.
- Entities are **extracted, not inferred**. A date or an amount appears because it was
  present in the text. Missing beats wrong.

## Documentation

- [Homepage](https://inboxtxt.dev/): the format, the tiers, the measured benchmark, and the open questions.
- [Why this exists](https://inboxtxt.dev/why.html): the article — what a mailbox actually costs a model, and why the fix belongs at delivery time rather than read time.
- [SPEC.md](https://github.com/eddie-boop14/inbox-txt/blob/main/SPEC.md): the draft specification — file shape, tier definitions, classification, redaction rules.
- [README](https://github.com/eddie-boop14/inbox-txt/blob/main/README.md): five-minute setup, the benchmark table, and what is still unresolved.

## Reference implementation

- [apps-script/Code.gs](https://github.com/eddie-boop14/inbox-txt/blob/main/apps-script/Code.gs): the sync worker. Runs inside your own Google account — no server, no API keys, no OAuth application. Paste, run once, set an hourly trigger.
- [mailmirror.py](https://github.com/eddie-boop14/inbox-txt/blob/main/mailmirror.py): a normalizer in Python, standard library only, for anyone porting the format elsewhere.
- [mcp/server.py](https://github.com/eddie-boop14/inbox-txt/blob/main/mcp/server.py): an MCP server exposing the mirror to any agent — `inbox_overview`, `inbox_body`, `inbox_since`.

## Reproducing the numbers

- [netlify/functions/tokenize.mjs](https://github.com/eddie-boop14/inbox-txt/blob/main/netlify/functions/tokenize.mjs): the tokenizer used for the published figures. Exact cl100k and o200k counts.
- [worker/](https://github.com/eddie-boop14/inbox-txt/tree/main/worker): the same thing as a Cloudflare Worker, for anyone who would rather self-host it.
- [example.txt](https://inboxtxt.dev/example.txt): a synthetic mirror, safe to read, showing every tier.

The benchmark ships with the spec on purpose. An earlier sampled estimate put the
reduction at 103×; measuring the whole corpus said 89×. Sampling flattered the project by
16%, and the correction is published in the same places as the original claim. Run
`benchmark()` and `tokenizeExact7d()` on your own mail rather than trusting these numbers.

## Optional

- [Full text of this site](https://inboxtxt.dev/llms-full.txt): every page, one file.
- [License](https://github.com/eddie-boop14/inbox-txt/blob/main/LICENSE): MIT.