← roster

Setu ★

R7 Eta · Systems
building

Middleware between a coding agent and the model provider. Fewer tokens, fewer wrong turns.

agent.setu { status : building ring : R7 Eta domain : Systems devices : setx users : - since : - publication : "Setu, positioning note" origin : square8 }

thesis

Coding agents burn tokens on three things that have nothing to do with the work: re-sending context that has not changed, guessing at underspecified instructions, and locking any saving inside one vendor’s product. Setu is not a better agent, it is the layer every agent passes through, which means adopting it does not require abandoning whatever you already use.

what it does

  • Intent clarification, converts silent failure into a cheap question
  • Deduplicates context across turns; caches stable context so it is not re-billed
  • Per-repo, per-developer, per-agent spend attribution
  • Tool-agnostic by design: Cursor, Claude Code, Copilot, Cline, Aider

what it does not claim

The obvious objection is that providers will make caching free. Partly true, and partly the point, provider-side caching only optimises within one provider. The defensible surface is what a provider will never build: cross-vendor neutrality, intent resolution before spend, and attribution an org can govern with. Three things must be true: savings clear a procurement threshold, latency stays invisible in an interactive loop, and compression never silently degrades output.

pieces

LLM gatewaycontext dedupintent resolverspend attribution

co-build

you want to work on this one with us. tell us what you'd take on.

i steal

not a request. tell us you're taking it & we'll get out of the way.