Codag

Codag is a compression and control layer for the tool outputs that coding agents read, targeting the repetitive lines that pile up from searches, tests, builds, logs, and API responses before native context compaction kicks in. The pitch to developers is direct: a single command attaches it to Claude Code and Codex, and it teaches the agent to read large tool results efficiently , so a multi-thousand-line test run or grep dump collapses to the failing lines and matches that actually drive the next decision, with the omitted bytes kept locally and retrievable on demand. The launch matters now because agent sessions increasingly stall on context bloat, where a single oversized tool result can evict what the agent already learned and force another expensive lap through the same commands. Codag positions itself as a drop-in proxy that leaves the model, prompts, and provider account untouched and only intervenes on large eligible results, classifying each one, reducing what is large, and exposing exact retrieval for the rest. For teams running Claude Code or Codex at scale, the appeal is cutting token spend and turn count without changing workflow or migrating tooling. Codag is a Y Combinator S26 company founded by Michael Zhou, who is running it as a solo founder out of San Francisco after previous engineering stints at Okta and Shopify.