Codag

Developers running coding agents such as Claude Code or Codex generate large volumes of tool output from test runs, searches, logs, file listings and API calls. The example given shows npm test printing 4,212 lines even though only a handful of lines about a failing test actually matter. Agents read this entire output as billed input tokens, and a single large result can push earlier context out of the window, forcing additional expensive calls. The page states that a session can make hundreds of tool calls, but only a handful carry the actual answer, meaning most of the billed input is read once and thrown away.

Codag installs with a single command, codag setup, which signs the user in, verifies the workspace with a live unbilled check, installs a local service, and attaches detected agents (Claude Code and Codex) without requiring new workflows or config files. Model traffic continues to flow through a local proxy directly to the user's own provider account, so the agent, model, prompts and credentials stay unchanged. Large eligible tool results (logs, test/build/lint output, search results, file listings, document reads, agent handoffs) are classified and reduced to the signal, such as failing tests or matching lines, while source code, diffs, configuration and structured data pass through unmodified. Omitted content is kept encrypted locally for seven days with a 1 GiB cap and can be retrieved exactly by line range, JSON path, or group when the agent needs it. The example shown claims about 75% compression on a test run, and the tool reports via codag savings and codag status how many tokens were avoided and sessions affected.