Skip to main content

Guides

Token Saver in Power Claude — spend less context on noise

Token Saver in Power Claude — spend less context on noise

Token Saver is Power Claude’s context optimizer: it trims low-signal noise from what the model re-reads (shell dumps, logs, diffs, JSON whitespace) so more of each turn is real work — not repeated tool spam.

TL;DR

ItemDetail
GoalFewer tokens per turn on repetitive tool output
DefaultOn (standard profile)
KnobsLevel / profile + per-type toggles (command, log, diff, json)
MeasureUsage tab · Token Tree · cost analytics
NotA free Max upgrade or rate-limit bypass

What it actually optimizes

ChannelTypical effect
Command outputCollapse huge ls / build / test dumps
LogsDrop heartbeats; keep errors/warnings
DiffsKeep changed lines; trim unchanged context
JSONLossless whitespace minify

You keep semantics that matter for coding; you lose bulk that only re-inflates the window.

How to set it

  1. Command: Power Claude: Token Saver (slider) or Settings → tokenSaver / contextOptimize.
  2. Start at standard / mid level.
  3. If answers get worse on huge refactors, step down to conservative.
  4. If you live in CI logs, keep log+command on.
  5. Pair with Warm Compacting so when you do near the wall, you don’t stall.

Prove it on your machine

  1. Open dashboard → Usage.
  2. Compare sessions with Token Saver on vs off (same task shape).
  3. Open Token Tree for a hot session — see which tools dominate.
  4. Savings shown locally are your numbers, not a brochure claim.

FAQ

Does it change my Anthropic bill formula?

It reduces tokens the model consumes in context. Plan limits and pricing still follow Anthropic. Power Claude’s subscription is separate.

Can it drop an error line I needed?

Log/command optimizers are designed to preserve errors/warnings; still verify on critical forensic sessions (turn level down).

Related

Primary sources

  • Features: Token Saver / contextOptimize settings in system index
  • Video: token-savings, token-tree scenes
  • Product page savings model (generated) on neural-llm.com/power-claude

Independent third-party software by Neural-LLM. Not affiliated with Anthropic.