AI Chat

You can chat in two places, and both share one history:

• Quick AI is a screen inside the palette. The search field becomes the message box, and replies appear above it as they stream in.
• AI Chat is a window, with your saved conversations in a sidebar on the left and the open one on the right.

You choose the model: one that runs on your Mac, a command-line tool you're already signed in to, or an API key of your own.

Turn it on in Settings → AI → Enable AI. It's off by default. While it's off, there are no Quick AI or AI Chat commands and no chat history file, and tab goes straight from the launcher to the clipboard.

Turning AI off stops any reply that's still streaming and closes the AI Chat window. It doesn't delete saved chats or API keys.

Opening a chat

For Quick AI:

• Run the Quick AI command in the launcher.
• Use its global shortcut. Press it again to hide the palette.
• Press tab from the launcher. Whatever you typed is sent right away as a new question.
• Choose the Quick AI row under "Use … with" at the bottom of any search. See Fallbacks.

For the AI Chat window:

• Run the AI Chat command, or use its own global shortcut. Both close the window again when it's in front.
• Press ⌘J in Quick AI. The conversation moves to the window, along with any text you've typed and any files you've attached.

Record both shortcuts in Settings → AI → Commands.

Quick AI

Type in the search field and press return to send. While a reply is streaming, return stops it.

The model name is on the right side of the header. Click it, or press ⌘P, to switch models. On models that support it, a second menu beside it changes the reasoning effort. A change applies to your next message and never interrupts a reply in progress.

Press esc in an empty message box to leave the chat. The conversation is still there when you come back.

Chat History

⌘K → Chat History lists saved conversations grouped by day, with a preview of the selected one.

A chat that's already open in the AI Chat window opens there. Delete All Chats keeps pinned chats. Empty chats are never saved.

The AI Chat window

The sidebar lists your saved chats, pinned ones first and the rest grouped by day. A new chat appears there once you send its first message. A spinner marks a chat whose reply is still streaming. Type in the field at the top of the sidebar to filter the list.

Type in the message box at the bottom. return sends, and ⇧return starts a new line. Below the text is one row of controls:

• + holds Attach Files…, Web Search (for models that can search), and Tools, which chooses the MCP servers this chat can use.
• The model and reasoning menus change the model for this chat.
• The ring shows how full the model's context is. Hover over it to see the tokens used, the cost where the provider reports it, and what your next message will send.
• The microphone appears while Dictation is on. Click it to dictate into the message box, then click it again or press return to insert the text.
• Send, which becomes Stop while a reply streams.

Your unsent text is kept when you close the window.

The Actions menu also has Copy Chat, Pin Chat and Delete Chat… once a chat is saved. Right-click a chat in the sidebar to pin, rename, copy, export as Markdown or delete it, or press delete to delete the selected one. If you're finding or renaming, esc ends that before it closes the window.

Each chat keeps its own model, so reopening a chat reopens its model too. Picking a model also makes it the default for new chats.

Replies

Replies are formatted as Markdown. Your own messages appear exactly as you typed them. A reply keeps streaming even if you close the palette or the window, or switch to another chat, and it's saved when it finishes.

Replies also render math. LaTeX between \( and \), or between single $ signs, appears inline; between \[ and \], or between $$ signs, it appears on its own line. An equation that's still streaming shows as … until it's complete, and copying an equation copies its LaTeX. Prices like "$5 and $10" stay as written.

Web pages a reply links to are listed under it as numbered sources, and each number also follows the sentence that cited it.

Suggested replies

When a reply ends with a few next steps, they appear as buttons. Click one to send it as your next message. To turn them off, use Settings → AI → Suggested replies, which is on by default. Apple Intelligence doesn't offer them, and they're also off while Send a system prompt is off.

Coming back to a chat

Two settings in Settings → AI control what you see when you open Quick AI or the AI Chat window again:

With Recent Conversation, you return to the chat you left, unless it's been idle for longer than the second setting. This also works after a restart.

Asking a question with tab or the fallback row always starts a new chat, so a new question is never added to an unrelated conversation.

Choosing a model

Models come from Settings → AI → Providers → Manage…. Default model, below it, sets which model a new chat uses and its reasoning effort.

The panel lists every provider on the left and shows the selected one on the right, in up to three tabs:

• Overview shows whether the provider is ready, which account it's signed in with, and which command it runs. The account's email address is blurred until you click it.
• Models lists every model the provider offers, each with a checkbox. Checked models appear in the model picker. If you haven't changed a provider's list, all of its models are shown, including ones it adds later. The default model is always shown.
• Advanced, for installed tools, is described below.

Every provider has a switch. Turning a provider off keeps it set up but removes its models from every picker, so you can set an API connection aside without deleting it or its key.

Apple Intelligence

Apple Intelligence runs on your Mac. It needs no key or account, and nothing leaves your Mac. When your Mac supports it, it's the default.

It works with text only, with no web search or attachments. If Apple Intelligence isn't available, Settings → AI says why. On a Mac that supports it, Open System Settings… takes you to its settings. It never switches you to a paid API without asking.

Installed AI: Codex, Claude, Grok, OpenCode and Cursor

If you already use the codex, claude, grok, opencode or agent (Cursor) command-line tools, Tinycast can use their existing setup. Codex also works with a configured API provider that doesn't require an OpenAI sign-in. OpenCode needs version 2 or later. Tinycast never asks for or stores their keys.

Each tool has its own switch, and all five are off by default. The pane shows whether each one is ready, not installed, or needs you to sign in. It links to the install page and can copy the sign-in command for you.

Tinycast finds each command the same way your Terminal does. If it finds the wrong copy, or none, open the tool's Advanced tab:

• Command path sets the command to run instead. Leave it empty to let Tinycast find it. If nothing exists at the path, Tinycast reports it rather than using another copy it finds.
• Variables are set for that tool every time it starts, such as a proxy or a config folder. Values are stored in your login Keychain. A few variable names are reserved by Tinycast to keep the tool from touching your files, and the row tells you when a value won't be used.

Tinycast uses these tools for chat only. Claude, Grok and OpenCode run with tools, file access and shell access turned off. The exception is MCP servers you add, which Codex and Claude can call but the other three can't.

Cursor runs in Ask mode in a private Tinycast workspace: it can read but not edit, and MCP tools aren't approved automatically. Cursor's command-line tool can't start without your MCP configuration, so MCP servers you've already approved in Cursor still apply. The Cursor row in the Providers pane points this out.

After each reply, Tinycast deletes the chat or session that the reply created. Your other saved chats in those tools aren't touched.

API connections

Use your own key with OpenAI API, Anthropic Claude, Google Gemini, OpenRouter, or any OpenAI Compatible endpoint, including a local one like Ollama.

• Keys are stored only in your login Keychain. They never appear in preferences, logs or backups.
• Remote endpoints must use HTTPS. Plain HTTP only works for localhost, 127.0.0.1 and ::1, where a key is optional.
• A key is tied to the address it was saved for. If you change a connection's base URL, Tinycast asks for a new key instead of sending the old one to the new address.
• While you edit a connection, Tinycast asks the provider which models your key can use, and you can search that list as you type. If a model isn't listed, or the provider can't list its models, type the model ID and choose Use "…" anyway.
• If a gateway needs extra headers, add them under Request Headers in the connection editor. They're sent with every request, replace a default header of the same name, and their values stay in your Keychain. With headers set, the API key is optional.

When you point an OpenAI-shaped connection at a gateway, meaning any base URL other than the preset's own, the reasoning effort menu offers Default and None. None turns thinking off on gateways that follow DeepSeek's API. Except on OpenRouter, each model in the connection editor also has a Reasoning levels menu. Check the levels the gateway accepts (Minimal, Low, Medium, High) and they appear in the reasoning effort menu too.

If a provider refuses a request, the reply shows the error with the provider's own message. Your key and header values are hidden from it.

Web search

Settings → AI → Web search is off by default. Your prompts are only sent to a search engine after you turn it on.

It works with Codex and with OpenRouter models. Searches appear inline in the reply with the query that was used, and sources are linked by name.

In the AI Chat window, + → Web Search turns the same setting on and off. While it's on, a Search button sits next to +. Click it to turn web search off.

Attachments

Press ⌘V in the message box to attach what's on your clipboard:

• An image or screenshot, scaled to at most 1568 px on its longest side.
• A PDF.
• A text file, like a CSV, Markdown or source file. Its contents are added to the message.

In the AI Chat window, you can also drop files on it or choose + → Attach Files….

In Quick AI, attachments appear as one small pill after your text, with an icon for the newest one and a count, like +2, for the rest. Hover over the pill to see their names, or click it to list them, with a small preview for each image, and remove any one with its ✕. Press delete in an empty message box to remove the last one.

Not every model accepts every kind of file. Tinycast refuses an attachment as soon as you add it and explains why, so it never sends something the model can't read.

Only files on your Mac are read. Copied web addresses are never downloaded.

System prompt

Tinycast sends a short note before every conversation that tells the model where it's running. The System prompt box adds your own instructions after that note. Apple Intelligence chats get neither, because any instruction steers the small on-device model's answers too much.

Send a system prompt (on by default) controls both. If you turn it off, no instructions are sent. Both are billed again with every message, and the pane mentions this.

When the box contains text, it opens blurred, so a screenshot of Settings doesn't show your instructions. Click the eye button to show and edit it.

Tools from MCP servers

With an API connection, or with the installed Codex or Claude command, the model can call tools from MCP servers you add. See MCP servers.

Privacy and storage

• Conversations are saved on your Mac in ai-chats.sqlite3, in Tinycast's Application Support folder.
• Keep conversations sets how long they're kept: 7 Days, 30 Days, 3 Months or Forever (default). Pinned chats are always kept. Old chats are only removed while AI is on, so they're kept on a Mac with AI turned off.
• No AI settings are included in backups: not the switch, the connections, the default model, the system prompt, web search, suggested replies or history settings.
• The signed-in Codex account is blurred in Settings until you click it.