Erepress

Overview

Every message you send to a model is paid for by the token. Some of those tokens carry no meaning: extra whitespace, padded formatting, and verbose machine-generated JSON the model reads the same whether it is dense or sparse. Erepress reformats that bloat away before the message is sent, so you spend fewer tokens for exactly the same result.

Erepress is lossless. It only deletes insignificant whitespace and compacts parseable JSON. It never summarizes, drops, reorders, or rewrites anything, so the model reads the same text it always would.

What It Does

Erepress applies a small set of safe, reversible reformatting steps to each message before it is sent:

• Compacts parseable JSON by removing the whitespace between its parts, keeping every value, key order, and character exactly as written.
• Trims trailing whitespace and collapses long runs of blank lines, while preserving Markdown hard line breaks.
• Leaves code blocks, secrets, and tool-call arguments untouched, byte for byte.

If a message cannot be reformatted safely, Erepress leaves it exactly as it was. It never makes a message larger, and it never changes meaning.

On by Default

Erepress is on for every workspace and every chat request, on every tier including the free shared-GPU tier. There is no plan to upgrade to, no setting to switch on, and no quality tradeoff to weigh: because it only removes insignificant whitespace, the model reads exactly what it would have read without it.

The trimming is deterministic and runs in-path. It adds no model call and no measurable latency to your request. Fewer tokens leave your request, so you are billed for fewer input tokens for the same result. That saving is yours by default; you do nothing to earn it. You can turn it off at any time, and the change takes effect on your next request.

How to Turn It Off

There are three ways to turn Erepress off, from the most local to the most durable:

• Per request. Send the request header X-Erepress: off. That single request is sent without any reformatting. This is the path for clients that need the bytes they sent delivered unchanged.
• Per workspace. The workspace owner can turn Erepress off for the whole workspace from the workspace Settings tab. The setting stays off until the owner turns it back on.
• Slash command. The owner can run /erebine erepress off to turn it off, /erebine erepress on to turn it back on, or /erebine erepress status to check the current setting.

A per-request header always wins over the workspace setting, so a single request can opt in or out regardless of the workspace default.

Guarantees

• Lossless. Only insignificant whitespace is removed. Values, key order, numbers, and string contents are preserved exactly.
• Never changes what the model reads. The model sees the same text it would have without Erepress, minus only the whitespace that carried no meaning.
• Never increases size. If reformatting would not shrink a message, the original is sent unchanged.
• Owner-controlled. Only the workspace owner can change the workspace setting.
• Reversible. Turn it off per request, per workspace, or with a slash command at any time.

Savings on Usage

The Usage page shows how many tokens Erepress saved over the window you are viewing, both as a series on the Token Usage chart labeled "erepress saved (est.)" and as an "erepress savings" value on the compact strip. The savings track the same rolling window as your other token, cost, and cache figures.

The savings figure is an estimate, computed from the characters removed (roughly four characters per token), and is always labeled "(est.)". It is illustrative only and is never the billed figure. Your bill always reflects the model's own count of the reformatted text.