Skip to content
MelMel
Reference > Settings & account

AI

Agent input, thinking display, voice, default models, and your provider keys.

The AI page configures how the agent behaves in conversations — input handling, the thinking stream, voice, desktop control, and which model new conversations start on. Changes apply immediately and persist to ~/.mel/settings.json.

Input

  • Autodetect terminal commands in agent input (on by default) — when you type something that looks like a shell command into an agent conversation, Mel runs it as a command inside the conversation instead of sending it to the model. Press Ctrl+I to override the detection for the current input, or prefix with ! to force a shell command. Turn this off if you want everything you type treated as a prompt.
  • Show input hint text (on by default) — controls the ghost autocomplete and the hint rows rendered under inputs.
  • AI command suggestions (on by default) — when your local autocomplete (session history, shell history, filesystem) has nothing to offer, Mel makes one lightweight backend call after a short typing pause and shows the suggested next command as ghost text. Disable this if you want autocomplete to stay fully local.

Thinking

  • Auto-open thoughts while streaming (on by default) — for models that emit reasoning, the collapsible Thoughts section expands automatically while the agent is thinking and collapses again when the response is done. With this off, Thoughts stays collapsed until you open it.

Voice

  • Voice input (on by default) — shows the mic on the input. The first recording asks before anything is sent; after that Mel's speech-to-text turns your words into text in the input, and the page shows how many minutes a day your plan includes. If you set a local command in $MEL_VOICE_CMD, Mel uses that instead and nothing is uploaded. See Voice input.

Desktop control

  • Let the agent see and control your screen (off by default) — when on, the agent can take a screenshot of your whole screen and use the mouse and keyboard on your real desktop. Only you can turn it on; the agent cannot. It works on Linux with X11 today and needs xdotool (or python3 with python-Xlib) and a screenshot tool (import, scrot or gnome-screenshot); on macOS and Windows the switch does nothing yet. Each action asks for approval like a command does — a screenshot included, listing window titles excepted — so a standing Always allow for commands, or autonomous mode, lets these actions run without a card. Another app that calls Mel never gets desktop control (see Use Mel from other apps).

Default models

Two dropdowns, each offering auto plus the models in the catalog:

  • Standard — new agent conversations start on this model.
  • Autonomous — the default for orchestrator and autonomous conversations, where the agent drives the terminal itself.

Both default to auto, which routes each task to a suitable model instantly.

  • Auto mode — which mode a conversation on auto starts in: Balanced, Cost-efficient, Genius or Open-weights. Picking a mode in the conversation's model menu sets it here too. See Model selection.

Provider API keys

Paste your own key for a model provider and the agent uses it for that provider's models, at no Mel credits. The key stays on this computer — in its keychain where there is one — and is sent only with your own requests. Signing out removes saved keys from this device. The avatar menu's Provider keys (BYOK) opens this section. See Models & pricing. Tip: Ctrl+Shift+click a model in the in-conversation model picker to select it *and* save it as your default here in one step.

Related