2026 hot

Token Diet

Unified installer and dashboard for RTK, tilth, Serena, and ICM — slashes AI session token costs by 40–90% without changing how you work.

Token Diet is an orchestration layer for four token-optimization tools that each attack a different part of the waste problem: RTK compresses CLI output before the agent reads it, tilth uses AST parsing to expose only the code symbols actually needed, Serena provides IDE-grade navigation so agents stop reading whole files to find one definition, and ICM persists decisions and resolved errors across sessions so agents stop re-deriving what they already know.

The stack installs globally once; all four tools operate with per-project context. token-diet health checks the full setup. token-diet dashboard opens a live browser view of savings across sessions.

What it does

  • RTK: compresses build output, test logs, and CLI noise by 60–90% before it hits the context window
  • tilth: AST-aware code reading — reads symbols not files; replaces grep, cat, find with structured equivalents
  • Serena: LSP-backed navigation and memory — finds definitions and references without reading whole files
  • ICM: cross-session, cross-tool memory — decisions and resolved errors persist across Claude, Codex, Gemini
  • token-diet gain shows cumulative savings; token-diet route <task> recommends which tool fits

Status

Cast. Running in the workshop as the token governance layer for all AI coding sessions.