Three third-party tools cut the tokens a Claude Code session spends, each at a different layer: rtk shrinks shell output, code-review-graph replaces searching and reading files, and headroom compresses whole requests on their way to the API. They do not overlap much, so they can run together.
| rtk | code-review-graph | headroom | |
|---|---|---|---|
| Layer | Output of shell commands | Finding code in the repo | Every API request |
| How it plugs in | PreToolUse hook on Bash | MCP server (+ hooks, skills, CLAUDE.md lines) | Local proxy via ANTHROPIC_BASE_URL |
| What it touches | git, test runners, grep, find, tsc… | Replaces grep + Read with graph queries | Tool output, JSON, logs, code in the newest turn |
| Misses | Built-in Read, Grep, Glob | Nothing, if the agent does not call it | Nothing on the request path |
| Getting the original back | Re-run unfiltered, rtk recall | Agent can still read files | headroom_retrieve |
| Claude Code features lost | None | None | Remote Control; tool search unless ENABLE_TOOL_SEARCH=true |
| Prompt cache | Kept | Kept | Kept (earlier turns untouched) |
| Telemetry | Off unless you opt in | None | On by default, HEADROOM_TELEMETRY=off |
| How to see savings | rtk gain | Your own transcripts | headroom savings |
| Author’s numbers (one month) | 27% overall, 79% in a test-heavy pet project | 3 calls in 18 sessions | 14% over 30 days |
The last row is one developer’s stats, not a benchmark. Each tool measures differently (rtk estimates tokens as bytes / 4; headroom says its own figure is likely high), so do not compare the percentages with each other.
When to pick which
- Start with rtk. It is one command to install, loses no features, and pays off at once in sessions with tests, diffs and logs.
- Add headroom when rtk is not enough, especially if much of your context comes from file reads and MCP results. Use the persistent install so tool search stays on, and keep a direct entry point (see Custom base URL) for Remote Control.
- Try code-review-graph on a mid-size or large repo, then check whether the agent actually calls it. If it rarely does, it costs little but saves little; the gain depends on the model choosing it over grep and
Read. - None of them replace the basics. Clearing between tasks, choosing the model and effort, and keeping CLAUDE.md short (see Reducing token usage) save more for free.
Related
- rtk, code-review-graph, headroom: each tool in detail.
- Usage limits: what these tools help you stretch.
- Reducing token usage: built-in habits first.