The compression engine for agent traffic

A pipeline of small, safe edits that shrinks every agent request before the provider bills it. It removes what the model has already seen and keeps what it needs.

See Anyray in action

Agents resend everything

Model APIs keep no memory, so every call carries the whole session again.

One agent request, late in a session
  • Stale tool outputOutput from steps the task has finished
  • RepeatsThe same file read twice, one header in nine files
  • NoisePassing tests, healthy log lines, JSON whitespace
  • Bulky formatsThe same field names on every record in a list
  • Needed nowThe new turn and the context it uses

Why naive compression costs more

Fewer tokens is not a smaller bill.

  • Broken cache. It rewrites old history on every call, so the provider bills that history again as new input. How we detect prompt-cache breaks.
  • Lost bytes. An edit tool fails on paraphrased code. The agent reads the file again, and the tokens come back.
  • Extra calls. A summarizer is a model call too, and you pay for it on every turn.

The saving that counts is the session bill for the same result.

Small edits, built for agent traffic

Exact bytes stay exact

Code, errors, the current turn and your users' words pass through unchanged. An edit tool always sees the real file.

Nothing is lost

Removed content goes to a store and leaves a marker in its place. If the agent needs it again, one call brings back the exact bytes.

It never breaks a request

On any error or timeout, the original request goes through as it was. A failure costs the saving on that one call, never a wrong answer.

Four kinds of waste, one pass

Each edit knows one kind of content, checks its own work, and stands aside when it does not apply.

  • Repeats

    A file read twice, one header in nine files. Later copies point to the first one.

  • Noisy output

    Passing tests and healthy log lines go. Failures and the lines that match the question stay.

  • Stale output

    Old tool output moves out of the request. The agent loads it back only when it needs it.

  • Bulky formats

    A list of records goes as one header row plus one row per record. Anyray checks that it unfolds back to the original first.

Model traffic still goes out raw

Databases have query planners. Networks have compression layers. Anyray is that layer for AI: it decides, call by call, what each request carries and what it costs.

Stop paying for the same tokens twice.