Context tokens sent to the model, with and without Hexum

Last updated 12 September 2026

A coding agent that is about to patch one function still has an incentive to read everything. Missing a file is a failed turn. Dumping the tree is a larger bill and a dumber prompt, but it rarely looks like a bug in the editor. That is why the default is eighteen thousand tokens for a three thousand token job.

Hexum sits on that path as MCP. The agent calls hexum_shrink before it reads widely. The model is sent this task, not every file. When it needs a full function it calls hexum_expand for that name. Hexum is not a new editor and not a compressor brand. Efficient is one of five products.

The published card

Study I, TR-2026-08-29. GPT-5.6 Sol. Official unit tests. Method: methodology. Interactive table: /benchmarks. CSV: results.csv.

Combined: 220,393 → 27,352 (−87.6%). 24/24 pass. Say 87.6% when you have 87.6%. Do not round it to 90%.

Estimator: ceil(utf8_bytes / 4). That is tokens sent to the model, not a cloud invoice and not Hexum’s bill. Hexum’s invoice is what the agent sent Hexum. Dashboard “tokens saved” is the cut. Details: what tokens saved means.

Tiny prompts save little

That is in the write-up on purpose. Hexum is for the turn where the agent was about to paste eighteen thousand tokens and three thousand would have passed the tests. Greenfield authoring cannot be shrunk. A focused agent that already opens four files can still beat a whole-repo shrink. Those cases are in Study II on /benchmarks.

Install

1. Get a key at the dashboard. First 2M tokens this month need no card.
2. Paste Hexum into Cursor or Claude Code.
3. Ask the agent to shrink this repo, then check architecture.hexum.

{
  "mcpServers": {
    "hexum": {
      "serverUrl": "https://hexum.dev/mcp",
      "headers": { "Authorization": "Bearer hexum_live_…" }
    }
  }
}

Compared with sending every file: repo dump vs Hexum. Architecture check that does not call a model: architecture.hexum.