A free, open-source CLI for Claude Code, Codex, opencode and your own API calls, and a VS Code extension. It reads the usage already on your machine, and every line of it is on GitHub under MIT.
It finds wasted tokens, prices every fix in exact dollars and applies the ones you accept. Guard holds a message to an expired cache: in my last 30 days, 40 of them wrote it again for $90.26, where a warm cache would have read it for $2.26.
Claude Code’s context, its cache countdown and the 5-hour window in VS Code’s status bar, and a heads-up 5 minutes before an expensive cache expires. On mine: reply before then to skip a $1.14 rewrite. The savings, guard and the fixes are one click away.
“good to see the per-model read rates in there, that’s the part most tools still have at a flat 0.1x.”
“I tried optimAIzr, it’s really good. Great work, thanks.”
One optimaizr profile run on Oct 25, 2026, over 30 days of my own Claude Code and Codex use. Every call re-reads the whole conversation. Printed below as the receipt my AI bill never came with.
optimaizr adds up every saving, none counted twice: up to 30% of my usage, about 1.4x the work per 5-hour window. And guard warns when your pace will run out the limit: it came before all 7 of my stops.$ optimaizr guard on0.11Oct 13, 2026Did the change work? MeasuredEach change you make is measured: cost per prompt in the days after it against the days before, with what the saving cost you. And live says what each turn cost.now 0.11.1 · opencode joins Claude Code and Codex: its sessions in every report and in optimaizr live.$ optimaizr resultsOn my machine, yes. 12 days after I started compacting at 200K, optimaizr results measured a prompt at 5% to 24% less than the 3 to 14 days before, and showed what it cost me: 6.8 minutes a day waiting for summaries. Tried on my last 51 five-hour windows, guard's pace warning came before all 7 of my stops at the limit, 67 minutes ahead on the median. Run it on your own history and see. If a number looks wrong, tell me.
No. Model choice is one of fourteen rules, and on most Claude Code bills it is not the biggest lever: every call re-reads the whole conversation, so compacting earlier usually saves more. The other rules catch files read twice, blind retries, oversized tool output, heavy reasoning on trivial answers and coming back after the cache expired.
Every saving it found, together, with no dollar counted twice: the biggest first, and each one after it only on what the ones before leave. A plain sum of mine would have said 33%. It judges whole tasks instead of single calls, and leaves out normal debugging, files being written and switches that would not pay back. Each part says how sure it is, and the biggest ones are an "up to" until you test them. On a plan, 30% less usage is about 1.4x the work in each 5-hour window and each week.
Your usage never does. It reads token counts from files your agents already write. verify calls your own provider with your own key, live --jev is opt-in, and once a day it asks npm for the latest version number, which you can turn off with OPTIMAIZR_NO_UPDATE_CHECK=1.
Yes. The CLI is MIT-licensed and developed in the open on GitHub. Pro, the hosted version, is a separate product.
The CLI is free, with no account. Pro, a hosted version for you or your team, is coming soon.
Claude Code, Codex and opencode sessions, your own Anthropic or OpenAI calls through wrap(), or a CSV export.
Yes. optimAIzr for VS Code puts Claude Code's context, cache countdown and 5-hour window in the status bar, with the savings, guard and the fixes in a sidebar. Search optimAIzr in Extensions. The CLI runs in VS Code's terminal too.
just trying it once? npx optimaizr profile runs without installing
A hosted version for you or your team: a dashboard, 12 months of history, auto-apply fix and spend guardrails, with Jev double-checking every model switch.
The CLI stays free either way.