MoeDocs
Start

Settings › Brain

Moe has no mind of its own; it borrows one you already have. Settings › Brain is whose, which model, and how hard it thinks: three things on one page, and no Apply button. Tapping a chip saves it, hands it to the part of Moe that thinks, and checks that Moe still answers.

The account card#

The card at the top names the account Moe is borrowing, Claude, on your monthly plan, Anthropic, with your key, Ollama or a local server, on this Mac, with a dot and one line saying how it is: Working, on Sonnet. (or Working, on Sonnet, high effort.) while all is well; Waiting for a sign-in., Not signed in yet. or Starting… while it is not; and, while a change goes through, Saving…, Switching to ChatGPT… or Checking Opus… until the verdict lands: It works. Moe is thinking with Claude (Opus)., or a plain sentence about what went wrong. If a sign-in is outstanding, the button for it is right there: Sign in with Claude… or Sign in with ChatGPT…. Under the name, one sentence on what this way of paying costs, Nothing on top of the plan you already have.

Change account…#

Change account… opens the one sheet with fields in it, Which account should Moe use?, listing every choice in three groups:

  • A plan you already pay for — Claude and ChatGPT, both marked Recommended, equally. Each signs you in through a page in your browser; for ChatGPT, Moe shows you the code the page asks for.
  • A key you have — you pay by the word — Anthropic, OpenAI, OpenRouter, Google Gemini, DeepSeek, xAI (Grok), and Anywhere else with your own key and address. Each takes Paste your key here; the key is kept on this Mac, readable only by you, and never shown again (a row that has one says Key stored). Anywhere else adds an Address field and a Test button that tries the key against it before you commit.
  • On this Mac — private, and free — LM Studio and Ollama or any local server. Each takes an Address, the one your server answers on, ending in /v1; http://127.0.0.1:1234/v1 for LM Studio and http://127.0.0.1:11434/v1 for Ollama are filled in for you. Load a model there first. A local model is slower and less sure of itself than one in a data centre; whether it can look at your screen or use your connections depends on which model you load, and Moe says what the one you picked can do.

The row in use is marked In use; one you have signed in to says Signed in. Use Claude (or whichever you picked) at the bottom applies it; Cancel leaves everything as it was.

Model#

Model is a row of chips, the provider's own names with two words under each saying what picking it gets you, and Other… at the end for typing a model name the row does not show.

  • On a Claude plan: Opus (Smartest, slower), Sonnet (Quick and capable, where a new Moe starts), Haiku (Fastest), Fable (Newest).
  • On a ChatGPT plan: the models your plan actually offers, read from your account when you sign in. A saved model your plan does not offer is said so plainly, Your ChatGPT plan doesn't offer that, so questions on it fail, rather than failing quietly.
  • With a key, the provider's current models, from the fastest and cheapest to the smartest; OpenRouter's chips are the popular models across vendors.
  • Anywhere else, LM Studio and Ollama have no chips: type the name of a model your server has loaded.

Under the chips, one line, Can see your screen ✓ · Can use tools ✓, says what the model you picked can do, with a note when it cannot: This model can't see images, so Moe won't attach your screen, or No tool calling: Moe can talk but not act.

How hard it thinks#

Another row of chips: Default (Let Claude decide) and then the provider's own levels, from none or low up to xhigh or max, each captioned (No thinking, fastest, Quick, Everyday, The usual choice, Careful, slower, Slowest, most thorough). Higher is slower and more thorough; Default, where every Moe starts, lets the provider choose. On Claude, changing it restarts Moe's thinking part, which takes a few seconds; on a key it takes effect on your next question; a local model that does not think ignores it. A model with no levels shows no row.

How long Moe may work on one thing#

On a Claude plan, a row of four chips: 5 minutes (Quick jobs only), 15 minutes (Most errands), 30 minutes (The usual choice, and where every Mac starts) and 1 hour (Long research). It is a ceiling for one request, not a timer on every answer: most questions are done in seconds whatever you pick. If a task is still going when time is up, Moe stops it and tells you, That was taking longer than Moe may work on one thing, so Moe stopped it, and you can give it longer here. Tapping a chip takes effect from your next question.

When the account Moe borrows stops working#

Two very different things can go wrong with the account Moe is borrowing, and Moe tells them apart.

Something that will pass. Your provider is busy, or limiting how fast you can ask. Moe says so where you asked, "I've hit the limit on your Claude plan — I'll answer as soon as it clears, in about 10 minutes", and then does exactly that. There is nothing for you to do.

Something that will not pass on its own. Your sign-in was cancelled, your key was revoked, the account ran out of credit, or the address Moe has for a local model stopped answering. Waiting will not fix any of these, so Moe never says it will. You are told where you asked, in the chat you typed into or out loud if you were talking, in your own provider's name: "OpenAI wouldn't let Moe in — the key is missing, wrong, or no longer good. Moe can't think until that's sorted: Settings › Brain." And it stays in Settings until it is fixed: the account card on this page carries an amber line, OpenAI turned down the key, followed by the one next thing to do, and Connections stops calling Moe's brain Working while it plainly is not. It clears itself the moment Moe answers anything at all; there is nothing to dismiss and no need to restart Moe.

A question your provider simply refused to answer, a safety filter, is said in the chat ("Gemini wouldn't answer that one. Try asking it a different way.") and kept out of Settings, because there is nothing about your account to fix. And nothing a provider says about your key ever reaches the screen: Moe strips anything key-shaped first.

If you have a cloud computer#

Changing your brain here changes your cloud computer's brain too. The moment you change it, a short line appears under the account card saying what became of your machine. Most of the time it says Your away machine is getting this too — it picks it up the next time it syncs, and uses it from then on. That covers a different model on the same account, a new key you just pasted, and your ChatGPT plan, where your machine is given a fresh copy of the sign-in on this Mac.

Three cases it cannot do by itself, and it says which. Your Claude plan: Anthropic gives each computer its own sign-in, so the line offers Sign your machine in too…, one click, approve in your browser, done; until then your machine keeps thinking with what it had. A key you pasted some time ago and did not retype: Moe will not send a key off this Mac that you did not just type, so the line offers Give your machine this too…. A model running on this Mac, LM Studio or Ollama: your machine cannot reach it, so nothing is sent and the line says why. If you have no cloud computer, none of this appears.

If the thinking part is not installed#

On a Mac that has the app but not yet the part of Moe that thinks, the page says so, Moe's thinking part isn't installed yet, and Set it up downloads it, about a gigabyte. Meetings and your connections work without it; answering a question is the part that needs it.