Use fewer tokens, losslessly. Aneirin is a local proxy that cuts the token cost of Claude Code and GitHub Copilot. It runs on your own machine, uses your own account, and does not change a single answer the model gives you.

Free to start — the full optimizer is free up to 5 million optimized tokens per day. Pro removes the cap ($5/mo or $79 lifetime). No sign-up to try it: download, run, save tokens.

In development on one machine, Aneirin has physically removed 652 million billed tokens from real requests so far. These are redundant tokens the model had already seen. Same answers, smaller bill.

Note

Lossless means the model still receives everything it needs to answer correctly. Aneirin only trims context that is redundant or re-billed, never the meaning of your request.

Download

⬇ Download the latest release — Windows today (macOS and Linux to follow).

Grab aneirin-windows-*.zip, unzip it, and follow the three steps below. Verify your download against the published SHA256SUMS if you wish.

Get started (Windows)

Aneirin is a single program — aneirin.exe. No Node or other install needed.

# 1. check it
.\aneirin.exe selftest          # → "Aneirin engine OK"

# 2. trust the local certificate (one time; generated on YOUR machine, never shipped)
.\aneirin.exe install-ca

# 3. run it (local optimizer on 127.0.0.1:8765 — leave it running)
.\aneirin.exe

Then send your AI tool's traffic through it:

# Claude Code (CLI) — set for the session, then use Claude Code as normal:
$env:HTTPS_PROXY = "http://127.0.0.1:8765"

That puts you on the Free tier (full optimizer, 5,000,000 optimized tokens/day). Watch your savings in the local dashboard / ledger inside your Aneirin folder.

First run shows a "Windows protected your PC / unknown publisher" prompt until the build is code-signed — click More info → Run anyway.

Full guide: START.md · what it does and the responsibilities: RISK.md · privacy: PRIVACY.md.

Buy a licence

→ Get Pro — paste the device code shown by aneirin.exe activate, pick a plan, and you get a licence token to paste back:

.\aneirin.exe activate "PASTE-YOUR-TOKEN-HERE"

The licence is verified locally and keeps working offline. It is bound to the one machine you bought it for.

Why Aneirin

• Lossless. Identical model behaviour. It removes re-billed context, not information.
• Local. Runs on 127.0.0.1. There is no Aneirin server in the request path.
• No telemetry. Your prompts, code, and usage never leave your machine. Free and Lifetime make zero network calls; Monthly refreshes only its own licence (device id only, once a day).
• Your own account. It only optimizes traffic you are already authorized to send.
• Zero workflow change. Keep using Claude Code and Copilot exactly as you do today.
• Fully reversible. One command uninstalls it. A kill-switch flips it to pure pass-through instantly.

How it works

flowchart LR
  A["Your editor<br/>Claude Code / Copilot"] -->|request| B["Aneirin<br/>local proxy · 127.0.0.1"]
  B --> C["Lossless trim:<br/>remove stale thinking<br/>de-dupe repeated context<br/>relocate cache markers"]
  C -->|slimmer request| D["AI provider"]
  D -->|same answer| A

Loading

1. Aneirin runs a small proxy on localhost. Your editor's AI traffic passes through it.
2. For each outgoing request it applies lossless reductions: it removes your own previous turns' "thinking" blocks the model no longer needs, de-duplicates byte-identical context, and places prompt-cache markers where the provider caches most effectively.
3. It forwards the slimmer request to the same provider endpoint your editor already used.
4. If anything goes wrong inside Aneirin, it forwards your original request untouched (fail-open). It never blocks or corrupts a request to "save" tokens.

The proof (real data, not estimates)

Measured across 6,699 real optimized requests on the developer's own machine: the typical optimized request is 17.5% smaller, with about 97,000 redundant tokens removed on average. The biggest single result so far removed 399,906 tokens of stale prior thinking from one call.

Aneirin counts only provably removed tokens (the ones it physically deleted from a request), never guesses. Its local dashboard shows your own saved total, and you can verify it against your provider's usage page. Nothing is uploaded to produce these numbers.

What "lossless" means (the guarantee)

Preserved exactly	Removed or relocated (never billed twice)
Your current question and instructions	Prior-turn "thinking" the model already consumed
Tool calls and their results	Byte-identical context repeated across turns
The frontier (latest) turn, in full	Cache markers moved to better boundaries

Losslessness is checked against a suite of real captured transcripts (including interleaved thinking and multi-turn tool use). Every transformation is asserted to leave the model-visible content byte-identical before it ships.

Important

Privacy. Aneirin contains no telemetry. Your prompts, code, traffic, and savings counts never leave your machine, and the licence is verified locally (it keeps working offline). Free and Lifetime make zero network calls of any kind. A Monthly subscription does one thing over the network: once a day it refreshes its own licence by sending only your device id (a random-looking hash), never your prompts, code, or usage.

Supported

Claude Code	Yes (VS Code and CLI)
GitHub Copilot	Yes (VS Code and CLI)
OS	Windows today. macOS and Linux to follow
Account	Your own. No shared or pooled access

Pricing

Aneirin is free to use — the Free tier runs the full optimizer every day, no account, no card. Upgrade to Pro only when you want the daily cap removed.

Tier	Price	What you get
Free (community)	$0 forever	The full optimizer, capped at 5 million optimized tokens per day (combined across projects). Dashboard and ledger always on. No account, no card.
Pro, Monthly	$5 / month	No daily cap. Cancel anytime.
Pro, Lifetime	$79 once	Everything in Pro, never expires, one machine.

For a token-heavy developer the lifetime licence typically pays for itself within about a month, and the dashboard shows you exactly how much you have saved.

FAQ

Does it change the answers I get? No. It removes context the model has already used or that is repeated byte-for-byte, and it relocates cache markers. The model's input meaning is unchanged, so output behaviour is the same.

Why is it closed-source? The optimizer is the product. In place of "read the code," Aneirin earns trust by running locally, making no network calls of its own, verifying its licence offline, and being fully reversible with a one-command uninstall.

Can I trust a proxy with my prompts? Your prompts and code never leave your machine. The proxy binds to 127.0.0.1 only and forwards to the same endpoint your editor already called. The only Aneirin network endpoint is the licence service, and the only thing ever sent to it is your device id (and only on a Monthly plan, to refresh that licence). Your prompts and code are never sent anywhere.

What if it breaks something? It fails open. On any internal error it forwards your original request untouched. You can also drop a kill-switch file to make it pure pass-through instantly, or uninstall in one command.

Is this allowed by my provider? You are optimizing your own authorized traffic with a local tool. A full plain-language note on what Aneirin does, and the responsibilities involved, ships alongside the product — see RISK.md.

Uninstall

.\aneirin.exe uninstall      # removes the local CA trust + restores normal traffic

Then delete the folder. Aneirin makes no system changes beyond the local certificate it removes here.

---

_{© Corbenic AI, Inc. Aneirin is provided as-is under its LICENSE. You are responsible for your own use of it under your provider's terms.}