ScroogeLLM — a jovial Victorian gentleman laughing beside a vault stacked with bags labelled TOKENS SAVED, under a gilded sign reading 'Why pay full price?'

Every token counts. Especially the ones we don't spend.

Old Scrooge runs your books. He sits quietly between your editor and your AI providers, snipping off all unnecessary tokens, catching lazy LLMs trying to cut corners, and never wasting the prompts you've already paid for. You see every penny he's saved in your status bar.

Coming Soon to the Marketplace

Where did all your money go?

Most developers haven't the foggiest what each AI call actually costs them. Scrooge tracks every penny — racking up the savings in no time.

Before the Penny Drops
Code completions $12.80
Chat sessions $18.40
Inline edits $8.60
Agent tasks $7.40

Daily total $47.20
After Old Scrooge Steps In
Prompt trimmed -$8.20
Cache caught it -$11.40
Cheaper model sufficed -$6.80
Free optimisations -$2.40

New daily total $18.40
Old Scrooge sitting on a pile of moneybags labelled TOKENS SAVED and UNSPENT API, inside a vault marked SCROOGE LLM SAVINGS BANK
61% saved
an honest day's haul for this frugal developer

Every penny-pinching feature you'll need

From bone-honest cost tracking to silent auto-optimisation — all running on your own machine, where no provider can peek.

Every penny, as it spends

Watch each LLM call ring up in real time. Per-request, per-session, and a running tally in your status bar — nothing hidden.

Free Forever

Trim the fat from every prompt

Redundant tokens snipped before the API ever sees them. Same quality answer, leaner bill at the end of the day.

Never pay twice for the same answer

Ask the same thing again? The cache serves it instantly. No round trip, no token spent. Scrooge approves.

Cheap model when cheap will do

Not every task needs the maestro. ScroogeLLM nudges you toward the smaller model whenever it can carry the tune.

Real names stay home

Deterministic fake names replace real PII before it leaves your machine. Same input, same fake, every time — nothing leaks.

Free Forever

A ledger your CFO will love

Every request logs raw cost vs. actual cost. Inspect it, export it, and use it to justify your tooling budget at the next review.

Free Forever

API keys, in the vault

Stored in your operating system's native keychain. Never plaintext, never transmitted. Your keys, your machine, full stop.

Free Forever

Free tier, always on

Visibility, PII protection, and audit logging are free forever. Paid features add deeper savings, but you'll never need to pay to see what you're spending.

Free Forever

A miserly little proxy that earns its keep

Scrooge runs as a quiet local proxy. Your AI tools talk to the Scrooge proxy, which then ships the work out to the most appropriate LLM provider. In between, he pinches every penny he can — and reports back, honest to the cent.

1

Move in

Install from the VS Code Marketplace. Point your LLM tools at localhost. That's the whole setup.

2

Pinch the pennies

Every request flows through. Prompts get trimmed, PII gets scrubbed, and the cache catches what it can. All locally.

3

Count the loot

See real-time costs in VS Code. Browse your audit trail. Watch the savings pile up, one request at a time.

Why call the maestro when the cellist will do?

Not every prompt needs the most expensive model in the orchestra. ScroogeLLM conducts the whole ensemble — Haiku here, Sonnet there, Opus only when the music truly demands it. The tip jar fills up either way.

Old Scrooge conducting an orchestra of vintage brass robots labelled HAIKU, SONNET, FLASH, MINI, OPUS PREMIUM TIER and LOCAL FREE, with a sheet of music titled TOKEN ECONOMY IN B♭ and a savings tip jar on the podium

Let the bots build your empire

Agentic coders spend tokens like Edwardian construction crews spend bricks. ScroogeLLM keeps a careful eye on the worksite, vetoes the wasteful jobs, and reminds the foreman-bot when a smaller crew will do. You sit on the balcony with the cigar.

Old Scrooge on a brass balcony surveying robot construction crews building an Edwardian skyscraper labelled MEGACORP, FLOOR 12 REFACTORED, FEATURE #47, with a SCROOGELLM hot-air balloon overhead

Your secrets stay close to home

ScroogeLLM is built for people who take data seriously. The design goal is simple: the miser keeps the books on your machine, not ours.

Local-first by default

The proxy is designed to bind to localhost (127.0.0.1). Any remote exposure would require an explicit opt-in — and a good reason.

Native keychain storage

API keys go into your OS's native keychain — macOS Keychain, Windows Credential Locker, or Linux Secret Service — rather than plaintext on disk.

Telemetry-light

No analytics, prompts, or API keys leave your machine by default. If diagnostics ever get added later, they'll be off until you choose to turn them on.

Deterministic PII scrubbing

When enabled, real names, emails, and identifiers can be replaced with stable fakes before requests reach your provider. Same input, same fake, every time.

For the privacy geeks, here are the juicy details…

Community Intelligence — the fleet-tuned routing and the “users saved $X” figure — needs aggregate data from many installs to mean anything. It is off by default and gates only those two features; the proxy, cost tracking, and free-tier optimizations never depend on it. When you do opt in, here is exactly what we engineer so the stored dataset is genuinely anonymous, not merely “de-identified”.

No stable identifier — not even a random one

There is no install id, account id, device fingerprint, MAC, or hostname in the payload. We deliberately omit even a random per-install UUID: a stable random id is still a pseudonymous identifier, which would let two rows be linked to the same machine across sessions and would drag the data back under personal-data retention limits. Batches carry at most a non-persisted, unmapped per-batch token that is never stored and never tied back to you.

Numbers and enums only — no field can carry your text

The payload schema has zero free-text fields. It is numeric and enum values: stage timings, token-count buckets, and a feature-flag bitset. There is physically no string field through which a prompt, file path, API key, or code fragment could transit — so none can, even by accident. The ingestion endpoint validates every payload against a strict allow-list and rejects unknown keys outright.

One chokepoint, IP dropped at the edge

Everything flows through a single Cloudflare Worker that is the sole place any data leaves your machine — one file to audit. It strips cf-connecting-ip on the way in, so your IP is never recorded. No exact timestamps are kept; dimensions are bucketed, not stored at full resolution.

k-anonymity against singling-out

Every row we retain or serve sits inside an aggregate of at least k = 20 within its stratum. Sampling plus that threshold suppress rare combinations — an exotic model paired with a tiny token bucket, say — that could otherwise pick out a single user at low population. This is what defends against singling-out, linkability, and inference, the three re-identification routes anonymisation guidance cares about.

Method grounded in free public guidance — the ICO’s Anonymisation code of practice and the EDPB/WP29 Opinion 05/2014 on Anonymisation Techniques. We keep a dated, written self-assessment in the repo and re-review it whenever the schema changes; indefinite retention holds only while every box on that checklist still passes. Separately, anything pseudonymous (e.g. /good / /bad feedback keyed by a hashed id) is not covered by this and stays storage-limited.

We eat our own dog food

With a smile, even. Every shipped feature gets tested on its harshest critic first — the developer himself. He dogfoods this proxy every working hour. Which is why our figures aren't just forecasts — they're the savings we've already seen.

A cheerful developer happily eating from a bowl labelled DOG FOOD with a spoon
ScroogeLLM mascot tipping his cigar beside a gilded vault marked TOKENS SAVED, under a sign reading 'Why pay full price?'

Stop paying full price.

ScroogeLLM is on its way to the VS Code Marketplace. Free tier forever — paid tier saves you more. Old Scrooge insists.

Notifications launching soon