blendbunjaku/optimaizr
blendbunjaku/optimaizr has 246 GitHub stars and 9 forks. GitHot measured +3 stars in 24 hours. Last checked 2026-11-15 23:35 UTC. Last push 2026-11-12.
Measured star growth
- in 24h
- +3
- in 7 days
- Update delayed
- in 15 days
- Update delayed
- in 1 month
- Update delayed
Net star changes over overlapping periods, not daily totals. Missing measurements are left blank. Stars measure attention, not software quality.
About
Spend fewer tokens on the same work. A local CLI that finds wasted tokens in Claude Code, Codex, opencode and your own API calls.
AIREADME.md
OriginaloptimAIzr
Spend fewer tokens on the same work.
In a long Claude Code session, most of what you pay for isn't Claude writing code. It's Claude re-reading your conversation on every call, cheap per token but paid again and again (41% of my last 30 days), and writing it into the cache again after a break, at up to twice the input price: 58 of those cost me $98.83 in the same 30 days, where a warm cache would have read it for $2.48. optimAIzr shows where your tokens go, holds the expensive message before it is paid, and fixes what it can in one command. For Claude Code, Codex and opencode, in a terminal or in VS Code.
npx optimaizr
Runs locally: nothing leaves your machine. No account, no upload, no
telemetry. It reads the logs your agents already keep (~/.claude, ~/.codex,
opencode's database), keeps what it learns in ~/.optimaizr/, and by default
never reads your prompts or Claude's answers, only token counts. Guard, once you
turn it on, keeps a short note from your last prompt for /clear, deleted once
read or after 30 minutes. More under Privacy below.
Your Claude Pro did $801.89/mo of work at API prices: 40x what you pay
Up to 30% of it can be saved: about 1.4x the work per 5-hour window and per
week, if you take every step below.
Where the 30% comes from
Compact earlier 18%
Cold cache returns 8%
Smaller default model 3%
6 smaller ones 1%
Each is taken from what the ones above leave, so nothing counts twice.
Where it goes
Re-reading the conversation 41% ████████░░░░░░░░░░░░
Loading context into cache 34% ███████░░░░░░░░░░░░░
Answers, code and thinking 25% █████░░░░░░░░░░░░░░░
In VS Code
Most of my Claude Code runs in VS Code, where the chat panel draws no status line. optimAIzr for VS Code puts the conversation's context, its cache countdown and your 5-hour window in the status bar, and gives a heads-up 5 minutes before an expensive cache expires: on mine, "Reply before then to skip a $1.14 rewrite." The savings, guard, compaction at 200K and every fix sit in a sidebar, one click each.
143K · 54:51 · 5h 56%
Search optimAIzr in Extensions, or:
code --install-extension optimAIzr.optimaizr-vscode
The reel is my own run on Oct 29, with compaction at 200K already applied, so its sidebar shows less left to save (up to 13%) than the first run above.
Catch the expensive moments
The cache countdown. Claude Code keeps a conversation cached for an hour. Come back later and your first message writes all of it again. One command puts a countdown under Claude Code's prompt, with what coming back will cost.
optimaizr statusline on
◉ optimAIzr · 161.0K context, $0.032/call to re-read · cache warm 12m, then $1.29 to write again · leaving? handoff note, then /clear · 5h 61%
Guard: the expensive message, held once. With optimaizr guard on, a
message to a long conversation whose cache has expired is held once, with what
sending it would cost: on a Claude plan, its share of your 5-hour limit first.
Mine, after 3 days away:
optimAIzr held this message once. The conversation has been idle 3 days and its
cache expired, so sending it writes all 162K tokens again: about 3% of your
5-hour limit ($1.30 at API prices, Opus 5.5).
It was not sent: copy it from Original prompt below. Send it again to go ahead,
or type /clear and send it there: the new conversation starts from a short note
of where this one stood.
On Sonnet 5.5 the same rewrite is about 2% ($0.65): /model sonnet, then send it again.
Send it again and it goes through. Over my last 30 days, 40 messages like this
one wrote their conversation into the cache again for $90.26, where a warm
cache would have read it for $2.26. Guard also stops Claude reading a big file
whole, or running a command that failed twice a third time unchanged, nudges
/compact before a conversation compacts in the middle of a task, and warns
once when your pace will run out the 5-hour limit before it resets. It uses
Claude Code's own hooks, no mod, and optimaizr guard lists what it held.
Live, beside your agent. optimaizr live keeps a row for each open
conversation: a spinner and the cost so far while Claude works, how long the
cache stays warm once it stops. It warns when context jumps or passes 200K, and
5 minutes before a cache expires. ? shows what it checked, n the next
step for the biggest finding, a applies it. After each turn it says what
that turn cost and how much of your 5-hour window it took, and --notify sends
a system notification when the terminal is out of sight.
Where long sessions turn expensive. optimaizr sessions opens with your
costliest conversations, by the title Claude Code gave them, and what each one
took of your 5-hour window (mine, with the titles changed):
Costliest conversations (subagents included)
cost of 5h worked conversation
$91.16 ~220% 1h 53m Refactor payment webhook handling
$89.81 ~217% 2h 21m Migrate auth to session tokens
$74.40 ~179% 2h 36m Build analytics dashboard filters
$71.99 ~174% 2h 25m Add search indexing pipeline
$62.39 ~150% 3h 37m Fix flaky checkout integration tests
The top 5 of 80 conversations, 48% of spend.
5h share at $0.415 per 1%, learned from 8 windows of your plan meter.
Then it finds the context size where re-reading takes over, what each cache write was for, which conversations shared each 5-hour window, and every time a cache expired before the next call, with who was waiting:
Cold cache returns
58 times a long conversation's cache had expired by its next call, and
all of it was written again: 12.4M tokens for $98.83, where a warm
cache would have read it for $2.48.
you were away 51 $93.72
Claude was waiting on a tool 7 $5.11
Fix it in one command
optimaizr apply context-compaction # Claude Code and Codex compact at 200K
optimaizr undo context-compaction # and back
optimaizr results # did it work?
results compares the days after a change with the days before it, on your
own usage, and marks every row that got worse, so you see what a saving cost.
Every finding says what happened, why it matters and what it saves, in dollars from your own logs. Savings are split into clear waste, likely, and "up to". The total takes each one from what the others leave, so no saving is counted twice: a plain sum of mine would have said 33%, not 30%.
Share your numbers
The numbers above are one person's history: mine. npx optimaizr card makes
an image of your last 30 days with totals only, no project names, paths or
prompts, to post wherever you compare notes. If you applied a fix, add
optimaizr results, even when it says the fix didn't work.
Inside Claude Code
The optional optimAIzr mod runs inside the session: the cost of each answer and
your 5-hour window as you work, Y in live moves the session you're in to a
cheaper model, and /optimaizr handoff writes a note to start fresh from.
⏺ optimaizr: this turn $0.18 · 2m 10s · 4 requests · ~0.4% of 5h
optimAIzr █████████░░░░░░░ 56% of 5h · resets 15:00
/plugin marketplace add blendbunjaku/optimaizr
/plugin install optimaizr@optimaizr
Commands
| Command | What it does |
|---|---|
optimaizr |
Where your tokens go, and what to fix first |
optimaizr live |
Watches every open conversation while you work |
optimaizr guard on |
Holds an expensive message once, before it goes |
optimaizr statusline on |
The cache countdown under Claude Code's prompt |
optimaizr sessions |
Where long conversations start to cost more |
optimaizr why |
Spend by model, project and kind of work |
optimaizr apply <rule> |
Applies a fix; undo <rule> takes it back |
optimaizr results |
Did it work? The days after against before |
optimaizr verify <rule> |
Replays your own requests on the cheaper option |
optimaizr card |
Your last 30 days as an image to post |
optimaizr help lists the rest, optimaizr help all every view.
Works with
Claude Code, Codex and opencode 1.x with no setup. opencode history is read
from ~/.local/share/opencode/opencode.db (under $XDG_DATA_HOME when that is
set) and needs Node 22.13 or later. History from opencode 2.0 isn't read yet.
Reports, live and the steps in apply cover all three;
apply context-compaction changes Claude Code and Codex only.
In VS Code, the extension's status bar, heads-up, guard and done alert watch Claude Code; its savings read all three.
Your Claude plan (Pro, Max, Team) and ChatGPT plan are detected. Your own app
with one line, optimaizr.wrap(new Anthropic()) or new OpenAI(), and CSV or
JSON usage exports. Node 20.11 or later.
Recent releases
- 0.13 (Oct 30, 2026): optimAIzr for VS Code. The cache countdown, context and 5-hour window in VS Code's status bar, a heads-up 5 minutes before a cache expires ($1.14 to skip on mine), and the savings, guard and every fix in a sidebar. Search optimAIzr in Extensions.
- 0.12 (Oct 20, 2026): guard. The message that would write an expired cache
all over again is held once, with its price (58 of those cost me $98.83 in my
last 30 days), and
/clearstarts fresh from a short note of where it stood.liveshows every open conversation with its cache countdown,resultsmarks every trade-off, and open models in opencode are priced. In 0.12.1: the profile opens with everything it found, together and never counted twice (up to 30% of my usage, about 1.4x the work per 5-hour window and per week), your 5-hour limit is learned from Claude Code's own transcripts, and guard warns when your pace will run it out. Tried on my history, it came before all 7 of my stops, 67 minutes ahead on the median. - 0.11 (Oct 13, 2026): did the change work?
optimaizr resultsmeasures it on your own usage, with what the saving cost you.livesays what each turn cost and how much of your 5-hour window it took, andsessionsranks your costliest conversations the same way: my top one took $91.16, 220% of a 5-hour window. In 0.11.1: opencode support. - 0.10 (Oct 7, 2026): a cache countdown under Claude Code's prompt, no mod
needed. In 0.10.3: spot conversations that start heavy, and see what each
cache write was for in
optimaizr sessions. - 0.9 (Oct 6, 2026): your plan detected, savings labelled by how sure they are, compacting earlier in one command.
- 0.8 (Oct 2, 2026): the Claude Code mod, with the cost of every answer and cheaper models mid-session.
Every release in full:
changelog or
optimaizr changelog.
Privacy
Everything runs on your machine and stays in ~/.optimaizr/. To spot repeated
work it also reads what each tool call was given (a file path, a command, the
first 200 characters of anything else) and keeps that in memory only, since
commands can hold secrets. A report you export names the file or command a
finding is about. Guard, when you turn it on, logs what it held in
~/.optimaizr/guard/ for 30 days, a command as a hash and its first 80
characters with secrets masked. A held message leaves a short note from your
last prompt and Claude's last reply, deleted once read or after 30 minutes.
optimaizr guard print shows the hooks before anything is written, and
"manageSettings": false in ~/.optimaizr/config.json keeps optimAIzr out of
Claude Code's and Codex's settings for good: it says what to add by hand.
The only network calls are verify (to your own provider, with your own key),
the optional Jev second opinion, and a daily version check against npm that
sends nothing about you, off with OPTIMAIZR_NO_UPDATE_CHECK=1.
SECURITY.md
lists every file read and written.
Hosted dashboard, coming soon
optimAIzr Pro will be a hosted dashboard for you or your team: 12 months of history, fixes applied for you and spend guardrails. It is opt-in and separate from this repository: nothing is sent unless you connect the CLI to it, and never your prompts or code. The CLI stays free and local either way. Get told when it opens.
More
- Docs and the CLI reference: every command, flag and figure explained
- Contributing:
npm install && npm run build && npm test - Report a bug
First presented on September 13, 2026, open source since September 30, 2026.

