The AI assistant

How it differs from other AI

• The key is yours. Folding Mirror sells no credits and relays nothing. You enter your own API key and requests go from your computer straight to the provider you chose.
• It can see the document you're writing and search the whole library. So “make this paragraph shorter” is a sentence it can act on, with no copying and pasting.
• It stops and asks before touching your files. Editing a file, running a command or reaching an outside tool all produce a card first, laid out line by line; it reaches disk when you click “Apply”. (Interface things — focus mode, body size — just happen: you are looking straight at them, and one sentence puts them back.)

Step one: an API key

An API key is a string a provider issues you, which is how your usage is billed. Folding Mirror doesn't hand out keys — you register with a provider, add some credit (most start at a few dollars), and get a string that usually begins with sk-.

Anthropic, OpenAI, Gemini, OpenRouter, DeepSeek, Moonshot, Zhipu and others — 22 providers are preset here. Pick one and the endpoint and protocol fill themselves in; the key is all you deal with.

1. Press ⌘, for settings and choose “Providers & models”

   The gear at the bottom of the sidebar opens the same place.

2. Choose a provider

   The service address and message format fill themselves in. If you use a proxy or your own deployment, choose Custom… at the bottom, and those two fields appear for you to fill in by hand.

3. Paste the API key in and click “Save”

   It reports that the key has been stored encrypted. It is kept by the macOS keychain on this machine, written into no plain-text config file, and never leaves the computer.

4. Click “Fetch models”

   Folding Mirror pulls the list of available models from the provider, and you pick one under “Model”. If you know the model ID you can type it; left empty, the provider's default is used.

5. Click “Test connection”

   “Connection OK” and you're done. If it errors, don't switch providers just yet — common questions lists the typical errors and what each one means.

Step two: say something

Press ⇧⌘A to pull the assistant panel out from the right (the quill at the far right of the toolbar does the same). Type in the box and press return. The edge of the panel can be dragged if it feels narrow.

A few sentences that pay off immediately:

• “Shorten the second paragraph of this piece, keeping the tone.”
• “Give me an outline of what's here so far.”
• “Turn this passage into a flowchart.”
• “Check where this claim comes from and give me three sources.”

Whatever document is open is handed over as context automatically (the message shows a “current file” mark). It also works with nothing open, where it is simply an assistant.

Edits go through you

When the assistant means business — editing a document, creating a file, running a command, reaching an outside tool — an approval card appears in the panel: which document, what is added, what is removed, line by line.

• Click Apply: the edit reaches disk (and a checkpoint of the document is taken first).
• Click Refuse: nothing happens, and the assistant knows it was refused and carries on from there.
• You don't have to decide now. Switch sessions, quit Folding Mirror, come back tomorrow — the card is where you left it.

If approving each step is too slow, the model menu at the foot of the panel can switch approval mode from manual to automatic, where edits are written as soon as they're made. Running manual for a while first, until you know its habits, is the better order.

Letting it work the app

Where a menu item lives, and what a switch is called, is not worth remembering.Just say it:

• “Turn on focus mode.” “Stop the typewriter scrolling.”
• “The text is too small, make it bigger.” “Use the serif font.”
• “Show me the source next to the page.” “Open the outline.” “Hide the sidebar.”
• “Open my November note.” “Save this.” “Open version history.” “Export it as a PDF.”

It reads the state the window is in before it acts, so “one size bigger” and “put it back” are sentences it can follow. What it touched is left in the panel as anInterface card.

Model and thinking effort

The pill at the foot of the panel shows the current model; opening it gives three settings for this turn:

• Model: the list fetched in settings. Each is marked for image input, thinking and tool calls — to have it edit documents for you, pick one that supports tool calls.
• Thinking effort: light, medium, deep or maximum. To the right is cleverer but slower and dearer; to the left is quicker. Medium is plenty for everyday revision.
• Approval mode: manual or automatic, as above.

Skills

A skill is a set of instructions written for the assistant: how a certain kind of job should be done. Seven come built in and all are on by default —

You don't select a skill — say it in plain words and the assistant decides whether this sentence calls for one. Settings → Skills switches them individually; turning off what you don't use saves a little prompt.

You can add your own: any folder with a SKILL.md at its root can be imported (Import skill) — your own writing standards, say, or your team's house style. An edited skill takes effect on the next message.

MCP: outside tools

MCP is a standard interface that lets the assistant use tools that aren't part of Folding Mirror — looking things up, reaching your database, calling some service. Three remote servers are built in, working out of the box with nothing to install and no extra key:

• Exa: web search and page reading; the “web search” skill runs on it.
• Excalidraw: generating sketch canvases; the “draw” skill runs on it.
• Mermaid: validating and rendering diagram syntax, which is why the “diagram” skill rarely produces one that won't open.

To connect your own, go to Settings → MCP → Add server. Two ways:

• Remote: a URL is enough; if it needs authentication, put Authorization and friends in the headers.
• Local: a command (npx -y some-package, say) that Folding Mirror starts as a process and talks to.

Not using MCP at all is perfectly fine. For writing, revising and outlining, a model plus Folding Mirror's own reading and writing is enough.

How full the context is

The panel always carries a usage indicator: how much of this turn is fresh input, how much was a cache hit, how much is the reply, and what percentage of the model's window that adds up to. The use of it is practical — in a long conversation, or after pasting a lot in, you can see it filling up rather than waiting for the model to start forgetting things.

When it's nearly full, start a new session (New session at the top of the panel). Window sizes are recognised automatically; for a model it can't place, you can type the size in settings.

Session history

Every conversation is kept. History at the top of the panel groups them by date or by file, and “only this file” narrows it further — so while writing something long, the earlier rounds about it are all at hand. Sessions can be renamed and deleted.

Cost and privacy

• Cost: Folding Mirror takes nothing and no cut; usage is settled between you and the provider, whose billing page is where the number lives. To spend less: a cheaper model, a lower thinking effort, and a new session when a conversation gets long.
• What leaves the machine: what you typed, and the document text the assistant read to do the job. With the assistant unused, Folding Mirror doesn't go online.
• Where the key is: encrypted by the macOS keychain on this machine. It never leaves, and it never appears in a plain-text file.