MCP servers
Give AI Chat tools from Model Context Protocol servers, remote or running on your Mac.
Model Context Protocol servers offer tools a model can call: read a folder, search an issue tracker, look something up. Add a server and AI Chat hands its tools to the model while you talk.
Settings → AI → MCP Servers → Enable MCP servers holds the switch. It ships off, and it only matters while AI Chat is on. While either is off, no server is contacted and no server process runs.
Adding a server
Add MCP Server opens the editor.
| Field | What goes in it |
|---|---|
| Name | Anything. It becomes the server's handle, like @github. |
| Connection | HTTP for a remote server, or Command for one that runs on your Mac. |
| URL | For HTTP. Remote servers must use HTTPS; plain HTTP only for localhost. |
| Header | For HTTP. A header name and value, like Authorization and Bearer …. |
| Command | For a local server, like npx. |
| Arguments | Like -y @modelcontextprotocol/server-filesystem ~/Desktop. |
| Environment | NAME=value, one per line, like GITHUB_TOKEN=…. |
| Trust | Ask Each Chat, Always Allow or Never Allow. |
Test Connection runs a real handshake and reports how many tools the server offers, so a typo shows up here instead of halfway through a conversation.
Header values and environment values are stored in your login Keychain, never in preferences, logs or backups. Removing a server deletes them.
A local server runs with your own user account, so only add commands you trust.
Using tools in a chat
Tools from every enabled server are offered to the model. When it calls one, a row shows inline in the reply, with a spinner while it runs. The reply carries on after it, and the row is saved with the chat.
Talking to one server
Start a message with a handle, like @filesystem list my desktop. Only that server's tools are
offered, and the handle is removed before the message is sent. A handle that matches no server is
sent exactly as you typed it.
Permission to run
With Ask Each Chat (the default), the first tool call in a conversation asks first:
- Always Allow remembers the choice for that server.
- Allow This Chat allows it for this conversation only.
- Don't Allow refuses just this call. esc does the same, and the next call asks again.
Never Allow on the server's row keeps the server without offering its tools. Only Settings can set that.
A refused or failed call is not an error. The model is told what happened and can answer without it.
Which models get tools
Only API connections are offered tools: OpenAI API, Anthropic Claude, Google Gemini, OpenRouter and OpenAI Compatible endpoints.
Apple Intelligence and the installed Codex, Claude, Grok, OpenCode and Cursor commands never get tools. For those, chat works exactly as it does without MCP.
Limits
- A reply can go through at most 10 rounds of tool calls. A model that only keeps calling tools has stopped answering, so the reply ends and says so.
- Each tool result, and all results in one reply together, are capped in size.
- Servers start when you use chat and stop after 10 idle minutes, or when Tinycast quits.
- Tinycast offers nothing back to a server. Requests from a server, like sampling, are declined.
Backups
Neither the switch nor your server list travels in a backup. A server list can run code on your Mac, so adding one has to be something you do yourself.