Token Optimization for Claude Code

πŸ—’οΈ Description

A short catalog of tools that cut token usage in Claude Code and LLM work in general. Source: a LinkedIn post aggregating 10 repos. Treat as a starting point β€” most are unproven, more community experiments than vetted tools. Shared logic: less context, more output β€” via filtering, sandboxing, compression, or structural indexes. Overlaps directly with Context Engineering and Harness Engineering.

  • Source: LinkedIn post (Mohammed Aboelez, 2026-04)

🧩 Tools (10)

  • Caveman Claude β€” forces Claude to reply in a β€œcaveman” style, cuts ~75% output tokens with no accuracy loss.
  • RTK (Rust Token Killer) β€” Rust proxy that filters terminal output, 60–90% reduction, no dependencies.
  • Code Review Graph β€” Tree-sitter graph, Claude reads only relevant nodes β€” claimed 49Γ— reduction on large monorepos.
  • Context Mode β€” sandboxes raw output (logs, GitHub) into SQLite instead of dumping into context; claimed 98% reduction.
  • Claude Token Optimizer β€” a bundle of setup prompts; 90% savings on documentation (11K β†’ 1.3K tokens).
  • Token Optimizer β€” detects β€œghost tokens” (invisible characters/duplications) eating the context.
  • Token Optimizer MCP β€” aggressive caching and compression for MCP tools; 95%+ reduction.
  • Claude Context (Zilliz) β€” hybrid vector search MCP, the entire codebase as context at 40% lower cost.
  • Claude Token Efficient β€” a CLAUDE.md template that enforces terseness, zero code changes.
  • Token Savior β€” symbol-based (AST) navigation, not file-based; 97% reduction + persistent memory.

🧩 Taxonomy (mine, organized)

  • Output compression β€” Caveman Claude, RTK
  • Context sandboxing β€” Context Mode, Token Optimizer MCP
  • Code graph / symbol index β€” Code Review Graph, Token Savior, Claude Context
  • Prompt / CLAUDE.md templates β€” Claude Token Optimizer, Claude Token Efficient
  • Hygiene β€” Token Optimizer (ghost tokens)

☘️ Reasoning for me

  • My Agentic Systems and Qamera AI Claude Code sessions regularly exceed 200k context β€” the most sensible bets are code graph (Token Savior, Claude Context) + CLAUDE.md hygiene (Claude Token Efficient).
  • Context Mode (SQLite sandbox) fits log and GitHub API ingest scenarios perfectly β€” slots into Context Engineering as another layer.
  • Worth verifying repos before installing β€” this is a LinkedIn list, not peer-reviewed.

πŸ“– Further reading


Template: knowledge_note_info