HYPE OR FACT? · CHECKED 5 NOVEMBER 2026
Does Token Optimizer really cut coding-agent context waste and save real dollars?
OUR VERDICT
Partly factOur original verdict: “The product is real; the dollar savings are mostly the author's estimates. Token Optimizer is an active, well-documented plugin (about 2.9k GitHub stars, v5.13.63 as of 1 Nov 2026) that hooks into Claude Code and other coding agents to compress re-reads, checkpoint before compaction and show a local dashboard. The headline ~$598 / 24.8M tokens saved figure is the author's own 30-day observational snapshot, and most of the dollar total is estimated model-routing, not metered cuts. A third-party X comparison of one Claude Code task put it at about −18% tokens, behind several lighter tools. We haven't tested it hands-on.”
THE CLAIM
“Token Optimizer "cuts the tokens your AI coding assistant wastes", covers the "other 75%" of context that Headroom/RTK miss, and reports about $597.72 in all-in API-rate-equivalent value and 24.80M tokens saved (~8% of spent + saved) on a 30-day snapshot ending 1 Nov 2026.”
What’s actually true
- What it is: a Python/plugin toolkit by Alex Greenshpun (PolyForm Noncommercial 1.0.0, with a separate small-business permission file). Install as a Claude Code marketplace plugin, or via scripts for Codex, OpenCode, OpenClaw, Hermes, Copilot, Cursor and others. Zero telemetry claimed; state is local SQLite.
- Repo facts (GitHub, checked 5 Nov 2026): ~2,882 stars, ~225 forks, latest release v5.13.63 (1 Nov 2026), homepage token-optimizer.dev, description 'Find the ghost tokens. Fix them. Survive compaction.'
- Author's savings snapshot (BENCHMARK.md, 30 days ending 1 Nov 2026): ~$597.72 priced API-rate-equivalent subtotal across 1,644 main sessions. Breakdown: logged actions $78.59 (metered), model routing $422.88 (entirely estimated), repeat reads $60.69 (modeled), other estimates $35.56. Token card: 24.80M saved (~8% of spent+saved), of which only 9.83M is metered. The author states pricing is incomplete and the full net dollar value is unavailable.
- Feature claims (author's README): delta re-reads ~20%, structure maps ~30% (up to 99% per file), bash compression ~10%, search ~15%, lean-output nudges estimated 10–15% and 'not counterfactually measured'. Claims Headroom/RTK only cover 15–25% of context.
- Third-party talk (last ~2 weeks on X): @Slonski_rt (3 Nov) ran one repo-review task at 284K tokens in plain Claude Code and reported token-optimizer (plugin) at −18% tokens, behind Token Savior (−43%), a CLAUDE.md rules file (−40%), Caveman (−38%) and token-optimizer-mcp (−23%), and flagged the noncommercial license. @tomcrawshaw01 published a Hermes cost walkthrough that the author thanked on 27 Oct. No relevant Hacker News threads showed up in Algolia search.
- Licence caveat: PolyForm Noncommercial — free for non-commercial use; commercial use needs a separate permission (README/LICENSE). Not a typical MIT/Apache open-source licence.
Sources
Cost
Free to install. Licence is PolyForm Noncommercial 1.0.0 (not MIT/Apache); commercial use needs a separate permission. You still pay your coding agent's model/API bill.
USE IT IF
Developers on Claude Code, Codex, OpenClaw and similar agents who want automatic context compression, compaction checkpoints and a local cost dashboard, and who can live with a noncommercial licence or have permission for work use.
SKIP IT IF
Teams that need a standard commercial open-source licence, independently audited savings claims, or a tool that does not install hooks into the agent. Also anyone expecting subscription-bill refunds: the dollar figures are API-rate equivalents, not Max-plan credits.
How we knowChecked on 5 November 2026. Verdicts: Fact = the claim holds up; Partly fact = real, but with caveats or overstated parts; Hype = the headline claim isn't supported. Products change quickly, so check the maker's site before you buy. Our methods & sources ↗