Context compression: the gateway shrinks oversized request text (ADR-79). Every model call the gateway forwards is scanned on the device: repeated log lines collapse with counts, JSON minifies, and anything still over 12,000 characters (gateway.compress.maxBlockChars) keeps its head and tail with a marker naming what was cut. Identical requests compress to identical bytes, so the exact-match cache stays correct. The full original stays on the machine only (never synced) for 7 days, retrievable through the context_expand MCP tool on any harness. Off per call with x-hnl-cache: bypass, off entirely with gateway.compress.enabled: false. The Usage page reports compressed requests and the estimated tokens and dollars held back per harness.

The home page hero adds a "For teams" link to pricing and a proof line counted live from the public price table (how many models spend tracking covers); the line stays hidden when the table is unreachable rather than showing a stale number.
The usage guide opens with a first task (bank a turn, then find it) before the reference, and the memory-banking guide shows the current hnl:index markers.

Flag notifications