Asked, answered, no accordion.

A macOS app that cuts what Claude Code and Codex cost you to run. It sits between your agent and the provider as a local compression layer, shrinking the context each request carries before that request is billed. Same agent, same answers, fewer tokens on the invoice.

Two things, both on your Mac. A cache-aligning stage keeps the start of your request byte-identical from turn to turn, so the provider's cache discount survives. Then a router detects what each piece of content is, JSON, source code, logs, diffs, search results, and hands it to the compressor built for that kind. Only the newest content is touched: your system prompt, your tool definitions and every earlier turn are left exactly as they are, and no message is ever dropped from history. How it works has the full pipeline.

Two screens. Overview shows what you saved, tokens removed before the request left your Mac counted separately from the ones your provider discounted for you, split by agent and by day over seven or thirty. Agents connects and disconnects each agent in one click, with a pause switch for when you want Compresso out of the way. Features covers both.

Compresso never collects them. The Authorization header is copied through byte for byte, never stored, and never logged, and no credential ever reaches our servers. It is read in one place, on your Mac, in memory, for one purpose: working out which Claude or ChatGPT plan you are on, so the app can offer you the tier that matches. Your Claude token is used to ask Anthropic's own account endpoint about your plan over the session your agent is already signed in with, at most once a day for each of those sessions; your ChatGPT plan is already inside the token, so the app reads it locally with no request at all. The plan we work out stays on your machine, with one exception: when you open checkout, the app tells us which tier to bill you for, because that is what a purchase is. That happens when the checkout page opens rather than when you pay, and no other request carries it. Compresso never touches the credential files your tools keep on disk. The narrow read is structurally bounded: the request type used for routing and counting still has no field that could hold a credential, and the one function that reads the token is private to the app's core, so the background helper cannot call it.

No. Optimization runs entirely on your Mac. Requests go from your agent, through a local loopback port, directly to Anthropic or OpenAI. Compresso's servers never see a prompt, a file, or a line of your code.

Aggregate counters only: tokens before, tokens after, tokens saved, request counts, per agent, per day, plus your account email and device registration. No prompts, no code, no file names, no payloads. The privacy policy lists every field.

Your tools keep working. If the optimizer is down, still starting, or anything else goes wrong, requests pass straight through to the provider unmodified, in under a second, with no dead port. Quitting the app doesn't break your agents either; a background helper keeps the port answering.

Yes. Subscription traffic has no per-token price, so Compresso reports room, not dollars, headroom against your plan's limits instead of an invented dollar figure. Token savings are measured the same way regardless of how you pay.

The token count is measured, not modelled: context that never left your machine was never billed, by anyone, at any rate. Dollar figures are estimates, marked with ≈, priced against your real cache mix, never a flat uncached rate. The two are never summed.

It sets the base URL in your Claude Code and Codex config files, the apps' own settings, nothing else. Every write gets a timestamped backup first. It never touches your shell profile, your PATH, or another app's data.

Up to 2 Macs per account.

Three tiers, each matched to a provider plan: Pro for Claude Pro or ChatGPT Plus, Max 5x for Claude Max 5x or ChatGPT Pro, and Max 20x for Claude Max 20x or ChatGPT Pro 20x. Every tier includes everything — up to 2 Macs and a 7-day free trial — and every tier is 50% off for its first 3 months, billed monthly or yearly. The pricing page has the current figures.

Compresso quietly steps out of the way: requests pass through unoptimized, your tools keep working, and nothing breaks. An expired subscription is a passthrough, never a broken toolchain. Subscribe from inside the app to turn optimization back on.

Claude Code and Codex today. The intercept is agent-agnostic at the protocol level; more agents are a product decision, not a rewrite.

Yes, Compresso runs the open-source headroom engine (Apache-2.0), fetched at first launch and run entirely on your machine. Downloads are integrity-checked; a checksum mismatch is a hard failure, never a shrug.