blendbunjaku/optimaizr
Spend fewer tokens on the same work. A local CLI that finds wasted tokens in Claude Code, Codex and your own API calls. - blendbunjaku/optimaizr
10 stars · 0 forks
README
optimAIzr
Spend fewer tokens on the same work. optimAIzr is a local CLI that finds wasted tokens in Claude Code, Codex and your own API calls, shows what each fix would save, and checks it on your own traffic before you switch. No account, no upload, no telemetry. It reads token counts, never your prompts. First presented on September 13, 2026, and open source since September 30, 2026. npm i -g optimaizr optimaizr profile Or try it once without installing: npx optimaizr profile. New in 0.8.0: inside Claude Code Claude Code now runs mods, plugins that work inside it. The optimAIzr mod puts the numbers where you work, lets Y in optimaizr live switch the session you are in, and shows what each switch saved. ⏺ optimaizr: this turn $0.18 · 4 requests · 5h 55% → 56%
optimAIzr █████████░░░░░░░ 56% of 5h · resets 15:00 ❯
Install it from a Claude Code session (2.1.287 or later): /plugin marketplace add blendbunjaku/optimaizr /plugin install optimaizr@optimaizr
The cost while Claude works. The spinner shows what the turn has cost so far and how full your 5-hour window is. The band above the prompt shows the window, how long it lasts at this pace and when it resets, from Claude Code's own limit meter. Each answer gets one line with its cost, its requests and how far the window moved. Switches in the session you are in. Press Y on a model swap in optimaizr live and the running session uses the cheaper model from its next request. No restart, no /model. Subagents switch at once; a long conversation waits until reloading it into the new model pays back. A reasoning-effort finding lowers effort the same way. With optimaizr live --auto, confident switches apply without asking. What a switch saved. A switched turn's line leads with the saving: the same tokens priced on the original model, less what they cost. The band keeps a running total. Back in one command. Harder task? /optimaizr off returns the session you are in to its own model and effort; /optimaizr on resumes. A guard against retry loops. When the same command fails twice in a row with nothing changed, the next identical attempt is held once and Claude is asked to change something first. /optimaizr prints the session's spend, what it saved, both plan windows and any switch. /optimaizr hud shows it all as gauges and a sparkline.
It reads usage figures and the commands Claude runs, never your prompts or file contents, and makes no network calls. The source is in mods/optimaizr. What it does See where your tokens go optimaizr profile reads the sessions already on your machine and puts spend, waste and the biggest fix on one screen. Spend $476.70 $452.56/month at the last 30 days' rate Calls 4,062 Tokens 743.4M 739.7M in / 3.7M out
Flagged calls 931 22.9% of calls Potential savings $44.59/mo $542.48/year
! Model mismatch
17% of your requests use a model whose capabilities exceed the detected workload requirements. 671 calls affected.
Find out why optimaizr why breaks spend down by provider, model, project and kind of work, each level as a share of the one above. $132.47 total across 1,807 requests
anthropic $132.47 100% 1807 calls -> sonnet-5 $99.47 75% 1640 calls -> kopshti-back $93.34 94% 1541 calls -> Mechanical $36.11 39% 746 calls -> Reasoning $27.52 29% 335 calls -> Generation $15.22 16% 141 calls
Catch waste while you work optimaizr live runs beside your agent and raises a fix the moment a pattern shows up. Press Y and your app switches from its next request, and so does Claude Code with the optimAIzr mod. Without it, Claude Code switches from its next session. ⚡ optimAIzr
This task looks suitable for a cheaper model.
Current: Sonnet 5 Suggested: Haiku 4.5 Observed cost: $0.202 / 26 calls
[Y] Apply optimization [N] Continue [D] Why?
Know what a fix is worth Every finding comes with its own arithmetic: what the traffic costs now, what it would cost after, and how far to trust the figure. --why prints the full calculation and every assumption. MEDIUM 441 mechanical calls ran on an over-specified model $5.76/mo est. | $70.12/yr | needs verification | model-selection
now $14.07 -> after $6.75 (441 calls, 12% of spend) confidence medium quality impact medium basis estimated
Prove it before you switch optimaizr verify replays your own recorded requests on the cheaper option and scores the answers against your quality bar, so you never trade cost for worse output. PASS 40 samples replayed
cost/call $0.0121 -> $0.0034 projected $61.40/mo saved this check cost you $0.38
Quality checks ok no-refusal candidate 40/40, baseline 40/40 ok tool-name-matches candidate 39/40, baseline 39/40
Stay under your limits On Claude Pro, Max or Team, --plan pro shows your 5-hour sessions and how much of each goes on waste. On ChatGPT plans it reads OpenAI's own meter from Codex. On a company budget, --budget 300 names the day the money runs out. ChatGPT Plus $20.00/mo what you pay API-equivalent $117.72/mo 5.9x what you pay 5-hour window 82% used · resets 01:49 · a full window ~$4.76 Weekly window 38% used · resets Sep 25, 22:49 Waste 9% fix it and you'd hit the limits ~10% later
Works with
Source Setup
Claude Code None. Reads ~/.claude/projects.
Codex None. Reads ~/.codex/sessions.
Your Anthropic/OpenAI app One line: optimaizr.wrap(new Anthropic())
A usage export (CSV, JSON) optimaizr import usage.csv
13 detectors cover oversized models, cache misses, bloated prompts, repeated file reads, retry loops, runaway reasoning, cost spikes and more. Commands
Command What it answers
optimaizr profile Where am I wasting the most?
optimaizr why Where does the money go?
optimaizr live What is wasting tokens right now?
optimaizr mod Is the Claude Code mod installed and running?
optimaizr recommend What can I change, ranked by saving?
optimaizr simulate <rule> What would the change save?
optimaizr verify <rule> Does the output still hold up on my traffic?
optimaizr apply <rule> What exactly do I change?
optimaizr report A shareable HTML report
optimaizr card Your last 30 days as an image to post
Run optimaizr --help for everything. Privacy Everything runs on your machine and stays in ~/.optimaizr/. The only network calls are verify (to your own provider, with your own key) and the optional Jev second opinion. SECURITY.md lists every file read and written. Documentation
Package README: the full usage guide CLI reference: every command, flag and figure explained Changelog optimaizr.com/docs
Development
Path Contents
apps/cli The optimaizr command, published to npm
packages/core The analysis engine: pricing, ingestion, rules, reports
packages/local The Node host: wrap(), the local ledger, live tailing, replay
mods/optimaizr The Claude Code mod, installed from this repo's marketplace
npm install && npm run build && npm test Requires Node 20.11+. See CONTRIBUTING. optimAIzr Pro, the hosted product coming to optimaizr.com, is separate from this repository. License MIT