Setting up the Copilot

The Copilot is the AI layer: live suggestion cards during the call and the report afterwards. It is off by default and Parrot works perfectly well as a recorder and transcriber without it. When you want it, flip it on in Settings → Copilot and pick who does the thinking.

Three ways to power it

• Claude (cloud). The sharpest option. You bring your own API key (Settings → API Keys), transcript text is sent to Anthropic during calls, audio never is. Costs money per call; Parrot shows you exactly how much on each meeting.
• Ollama (local). Free and fully private, the AI runs on your Mac. Install Ollama yourself, then pick a model in Parrot's settings. If the model isn't downloaded yet, Parrot offers to pull it with a progress bar. Honest expectation-setting: local cards arrive noticeably later than cloud ones, because the model shares your Mac with live transcription. Great for reports; on live cards, expect a relaxed rhythm rather than instant. On Apple Silicon get the native build or it will be painfully slow.
• Custom server. Any OpenAI-compatible endpoint: your own machine, a company gateway, OpenRouter and friends. Paste the base URL and model name. If you're on a free tier with strict limits, the pace settings in the next topic are made for you.

Instant answers from your documents

On Claude, there is one optional extra: a TypeSafe AI key (Settings → API Keys). With it, the moment the other side asks something your documents cover, Parrot shows the matching excerpt as a "From your docs" card, usually within a second, while Claude is still writing its answer. Claude's card then takes over. Without the key nothing changes. It sends the question, a couple of lines of context and the matching snippets of your documents to TypeSafe, never audio, and it only works when your documents are tagged into the call profile you are using.

Different brains for different jobs

Live cards need to be quick and sharp. The post-call report has all the time in the world. So Parrot lets you pick a different backend for each: a common combo is Claude for the live cards and a local model for reports, which keeps reports free and private without slowing the call down. If you only set up one, both jobs use it.