Questions, answered.

Cursor, Claude Code and Codex, in their VS Code panels and in the terminal, with more tools on the way. It is on as soon as you install, and one license covers every tool on every Mac you sign in on.

macOS 11 or later on Apple silicon or Intel; Cursor, Claude Code or Codex (any of them, in VS Code or the terminal); about 1 GB of disk for the local models; and an OptiFlume account (Google, GitHub, Apple or an email link). No API keys.

Two small AI models that run on your Mac, about 1 GB together: a text-embedding model that helps judge each prompt, and a small language model that writes your working-memory notes. They download once after install and then work offline; nothing you type is sent anywhere to use them. They use the processor only for a moment, when you send a prompt or a note is written. In Claude Code and Codex each open session keeps its own copy in memory for now, about 1 GB per session; one shared copy for every session is in the works.

Only Cursor needs it: it lets OptiFlume finish an automatic switch in Cursor and send your prompt on the new model. It is optional; without it, OptiFlume suggests the model in Cursor and you switch yourself. Claude Code and Codex never need it.

macOS asks for your password once to install any installer package. The one-line Terminal install on the install page runs the same macOS installer, so it asks once too (in Terminal, with sudo). Updates never ask: they install in your own user folder.

Only models you can use: in Cursor, the ones turned on in Cursor's settings; in Claude Code and Codex, the ones your Claude or ChatGPT account can run.

Run optiflume harness off for the VS Code panels, or optiflume terminal off for claude and codex in the terminal; on turns them back on. Your editor settings go back exactly as they were. The installer's Customize page has the same two choices.

OptiFlume adds one hook to Codex so it can show why it switched. Approve it once in Codex, Settings, Hooks. If you use Codex only in the terminal, type /hooks in Codex, select OptiFlume's hook and trust it. One approval covers the panel and the terminal. Switching works without it.

You start claude and codex as before. In an interactive terminal session OptiFlume checks each prompt; a --model you pass sets the model you start on, and every prompt is still checked from there. Headless and scripted runs, such as claude -p or codex exec, and the tools' own commands, such as claude update or codex login, run with no switching. claude -p runs exactly as it would without OptiFlume; in codex exec, the Guardian and working memory still run once you have approved OptiFlume's hooks. Claude Code and Codex keep updating themselves as before.

Routing runs entirely on your machine and works offline. A Pro or trial license lets a machine work offline for about 30 days, plus a 14-day grace period.

Run optiflume uninstall. It removes the hooks, the extensions, the commands and the PATH line, puts your Claude Code and Codex settings back as they were, and keeps your settings and logs; optiflume uninstall --purge removes those too.

Auto chooses in Cursor's cloud, doesn't say why it picked a model, and bills each request at that model's list price; on Teams and Enterprise plans the model it picked is hidden by default. OptiFlume decides on your machine, from the code you are working on, shows every switch in the model picker with a note saying why, and lets you overrule any prompt. It only uses the models you have turned on in Cursor.

No. OptiFlume decides on your machine before anything is sent, so the prompt goes out once, on the chosen model. When it doesn't switch, nothing changes at all.

Not always. Good code comes from good context as much as from an expensive model. OptiFlume judges how complex each task is and sends it to the model it needs: hard work to the strongest model, routine work to a faster one. After your tool compacts a long conversation, the next model gets a short working memory of your goals and decisions (Pro), so it starts with the context it needs. When it is unsure, nothing changes.

A switch can cost a fresh cache write on the new model. OptiFlume only switches when the price gap is worth it, and after your tool compacts a long conversation it gives the next model a short working memory of your goals, decisions and failed approaches (Pro), so a switch does not start from scratch.

It decides in front of you and gives you the last word. Every switch is visible where you work, with a note saying why, and you can always choose not to route a prompt and keep your own model. You can pause it for a project, or altogether, at any time.

It fails open. If any part of OptiFlume errors, your prompt goes through unchanged on the model you selected. It cannot lock you out of your editor. Uninstall is one command.

It keeps a short working memory of each session as you go: what you are trying to do, the decisions made along the way, and the approaches that already failed. After your tool compacts a long conversation, the model gets that working memory, and so does any model OptiFlume switches to later in the session, so it picks up where you were instead of retrying what did not work. It works in Cursor, Claude Code and Codex, with nothing to set up, on Pro and during the trial.

On your Mac only, and it is never sent anywhere. The notes are in ~/.optiflume/working_memory, and the turn output they are built from is in ~/.optiflume/turns for 30 days. You can delete either folder at any time; the docs page on working memory shows how.

No. Sign in and the 30 days start the first time a Mac connects. The trial never turns into a paid plan by itself.

Automatic switching pauses and the free tier continues: tips on the right model, keeping your own model for any prompt, and the Guardian with its session watchdog. Subscribe any time to switch again.

The free trial is once per person, across Google, GitHub, Apple and email sign-in.

Every Mac you sign in on, on one account and one subscription.

Yes. The Guardian works on every plan, including the free tier, for as long as you use OptiFlume. Its secret and command checks keep running even while OptiFlume is paused; the session watchdog runs whenever OptiFlume is on. Sign in once on the Mac: while nobody is signed in, OptiFlume does nothing at all, the Guardian included.

Allowance, not dollars. Claude and ChatGPT subscriptions cost the same each month, but they come with usage limits that reset on a schedule, and larger models and higher effort use those limits up faster. OptiFlume moves routine prompts to a lighter model or a lower effort, so your allowance lasts longer and you are less likely to hit a limit in the middle of the hard work you want the strongest model for. Your console counts the switches; dollar savings are shown for Cursor, where OptiFlume can read real token counts.

No. One plan covers Cursor, Claude Code and Codex, in VS Code and the terminal.

Yes. macOS 11 or later, on Apple silicon or Intel; the download picks the right installer.

In your account console: Billing, then Manage subscription. Pro continues until the end of the period you paid for, then the free tier.

Ask us and we refund the whole months you have not started; Pro ends when the refund is made. Or keep Pro to the end of the year, with no further charges.

Yes. If you subscribe at $6.99 a month or $69 a year, that price stays for as long as the subscription does.

No. Everything OptiFlume reads and decides happens on your machine, and OptiFlume never sends your prompts or code anywhere. (Cursor, Claude Code and Codex still send your prompt to the model you use, as they always have.) The routing engine works offline.

Sign-in and a plan check, an update check, a signed compatibility list, and a daily rollup of counts for your console (turn it off with optiflume telemetry off). Never your prompts or code. The full list is in What is stored where.

No. Cursor, Claude Code and Codex keep your model and account credentials. OptiFlume changes the model through their own controls, never reads or changes your Claude Code or Codex memory files, transcripts or sign-ins, and stores only its own sign-in session.

On your machine, under ~/.optiflume, in plain text and SQLite you can open. Our servers keep your account, your plan and, if you leave it on, daily usage counts.