AI Chat

AI Chat is a conversation screen inside the palette. The search field becomes the message box, and replies stream in above it. You choose the model: one that runs on your Mac, a command-line tool you already signed in to, or an API key you bring.

Settings → AI → Enable AI holds the switch. It ships off. While it is off there is no AI Chat command, no chat history file, and tab skips straight from the launcher to the clipboard.

Turning AI off stops a reply that is still streaming. It does not delete saved chats or API keys.

Opening a chat

• The AI Chat command in the launcher.
• Its own global shortcut, recorded in Settings → AI.
• tab from the launcher. Whatever you typed is sent straight away as a new question.
• The AI Chat row under "Use … with" at the bottom of any search. See Fallbacks.

Talking to it

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

Replies render Markdown. Your own messages stay exactly as you typed them. A reply keeps streaming even if you close the palette or open another screen, and it is saved when it finishes.

The model name sits at the right of the header. Click it to switch models, or to change the reasoning effort on models that support one. A switch applies to your next message; it never interrupts a reply already on its way.

esc on an empty message box leaves 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 one selected.

Empty chats are never saved.

Coming back to a chat

Two settings in Settings → AI decide what you see when you open AI Chat again:

With Recent Conversation, you land back in the chat you left, unless it has been idle longer than the second setting. This works across a restart too.

Asking a question with tab or the fallback row always starts a new chat. A question asked outright should not be tacked onto an unrelated conversation.

Choosing a model

Settings → AI → Providers → Manage… is where models come from. Default model below it picks the one chat uses, and its reasoning effort.

Apple Intelligence

Runs on your Mac. No key, no account, and nothing leaves the machine. When your Mac supports it, it is the default.

It reads and writes text only, with no web search and no attachments. If Apple Intelligence is switched off in System Settings, Tinycast tells you so. It never quietly moves you to a paid API instead.

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 them with the account you are signed in to. Tinycast never asks for or stores their keys.

Each one has its own switch, and all five ship off. The pane shows whether each is ready, missing, or needs you to sign in. It links to the install page and can copy the sign-in command for you.

Tinycast uses them as plain chat. Claude, Grok and OpenCode run with tools, file access and shell access switched off. Cursor runs in Ask mode against Tinycast's private workspace: read-only exploration, no edits, and no MCP auto-approval. Cursor's CLI has no way to start without your MCP configuration, so MCP servers you have already approved in Cursor still apply on this route — the Providers pane says so on the Cursor row. After each reply, Tinycast deletes the chat or session that turn created — not your other saved chats.

API connections

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

• Keys live only in your login Keychain. They never appear in preferences, logs or backups.
• Remote endpoints must use HTTPS. Plain HTTP works only for localhost, 127.0.0.1 and ::1, where a key is optional.
• A key belongs to the address it was saved for. Point a connection at a different base URL and Tinycast asks for a new key rather than sending the old one somewhere new.
• While you edit a connection, Tinycast asks the provider which models your key can use, and you search that list as you type. If a gateway cannot list models, type the model ID by hand.

For a gateway that copies DeepSeek's API, the reasoning effort menu offers None, which turns thinking off.

Web search

Settings → AI → Web search is off by default. Your prompts only reach a search engine once you turn it on.

It works with Codex and with OpenRouter models. A search shows inline in the reply, with the query it used, and sources are linked by name.

Attachments

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

• An image or screenshot. Scaled to at most 1568 px on its long edge.
• A PDF.
• A text file, like a CSV, Markdown or source file. Its contents go into the message.

Each attachment shows as a small pill beside what you type, with its own ✕. After two pills the rest fold into a +N count. delete in an empty message box removes the last one.

Not every model can take every kind. Tinycast refuses at the moment you attach, and says why, rather than sending something the model will never see.

Only files on your Mac are read. A copied web address is never downloaded.

System prompt

Tinycast sends a short note ahead of every conversation that tells the model where it is running. The System prompt box adds your own instructions after it.

Send a system prompt (on by default) controls both. Turn it off and no instructions are sent at all. Both are billed again on every message, which is why the pane says so.

The box opens blurred when it has text in it, so a screenshot of Settings does not show your instructions. Click to edit.

Tools from MCP servers

With an API connection, 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 decides how long: 7 Days, 30 Days, 3 Months or Forever (default). Old chats are only removed while AI is on, so a Mac with AI switched off keeps them.
• No AI setting travels in a backup: not the switch, the connections, the default model, the system prompt, web search or history rules.
• The signed-in Codex account is blurred in Settings until you click it.