Skip to main content
Raven calls your Anthropic or OpenAI key from the desktop app. There is no Raven AI proxy.

​
Providers and models

Configure Settings → API Keys (and the model / effort pickers). Examples from the in-app catalog (the list in Settings is the source of truth): Haiku or Luna are a good default for live Assist. Pick effort (low → max) when the model supports it.

​
Overlay actions

Cmd/Ctrl + Enter runs Assist and includes a screenshot when requested. You can also type a question in the overlay input. Mode instructions (topics, tone, canned replies) are treated as instructions, not optional flavor.

​
Session context (not long-term memory)

Each Assist call is a new API request. For a long recording Raven sends:
  1. A compressed session memory (cheap model, in RAM)
  2. Pinned opening transcript and your typed questions
  3. The last few turns
  4. The live transcript tail
That memory dies when the session ends. There is no cross-meeting user profile. History you reopen later lives in SQLite, not in that RAM structure.

​
RAG

If the active mode has documents, Assist retrieves the top MiniLM matches and injects them into the system prompt. See Modes.

​
Screenshots

Cmd/Ctrl + Enter can attach a PNG from desktopCapturer. Raven tries to keep its own overlay out of that frame.
  • During a recording: screenshot + transcript + mode + RAG
  • Not recording: visual content only

​
Formatting

Responses render markdown, KaTeX math, and highlighted code.

​
Copy

Hover a response card and use the copy control to put it on the clipboard.