One panel for the conversation you're in.
Press ⌘⌥H and a floating panel joins your call. It hears both sides through macOS itself - no bot joins the meeting, no account exists, and with the built-in model nothing is uploaded.
Included with every plan · Free 14-day trial · macOS 26 or later
Assist - answers while you're still talking
Both sides of the conversation are transcribed live. When a question comes your way, a suggested answer appears on its own - or when you press ⌥⌘A from any app, so you're never seen clicking into a panel on camera.
Answer one specific line
Hover any line in the transcript and answer exactly that - either speaker, however far back it was said.
One-press rewrites
Make an answer Shorter or Simpler, or ask for More detail or an Example - one click instead of retyping the question.
Templates for the situation
Save the scenario, who you are and any background - for interviews, sales calls or standups - and reuse it.
Ask about what's on screen
Capture a still of one display, excluding Dictera's own windows, and send it with your next question. Needs a model that accepts images.
Meeting Notes - the write-up, without a bot
Same capture, different result. Press Stop and Dictera writes the call up: summary, decisions, action items, open questions. Nothing joins your call and nothing is uploaded when you use the built-in model.
Four levels of detail
Notes, full minutes, a cleaned-up transcript in the language it was spoken, or chapters - a timeline of what was discussed when.
Shaped to the session
A lecture has no decisions and no owners; a meeting does. Dictera reads each recording for what it is and says which shape it used.
Action items with owners
Name the participants and an item can say who owns it - and "Unassigned" where the meeting never established who, rather than a guess.
Long meetings survive
A transcript longer than the model's window is read in passes, and the result says so when it covers only the most recent stretch.
Chat - a private model, one shortcut away
The third mode captures no audio at all. It's a typed conversation with whichever model you've pointed it at, including the one already running on your Mac. No browser tab, no account.
Nothing is listening
Chat never opens the microphone or the system-audio tap.
Follows the language you type in
Unlike the listening modes, Chat answers in whatever language you write.
Its own engine
Point Chat at the built-in model, a local server or a cloud provider, independently of the other two modes.
Screenshots work here too
Capture a display and ask about what's on it, with the same viewer and the same rewrites.
How it hears the room
Two audio streams, transcribed separately and merged into one two-speaker transcript - you on one side, everyone else on the other.
No Screen Recording permission
The other participants are captured with a Core Audio process tap, so a meeting needs only microphone and audio-recording access - and the call keeps playing normally.
Your mic starts muted
Every mode begins with your own voice off. Unmute when you want to be in the transcript; the panel always shows which state you're in.
Speaker echo, cancelled
On loudspeakers the other side would be transcribed twice - once from the call, once as room echo. Dictera gates the mic while they're audible, and stands down when you're on headphones.
Pick the device you actually hear on
Capture the participants from your AirPods, or leave it on System Default to follow whatever the output is.
Each mode picks its own model
Something fast and local for live suggestions; something larger for the end-of-meeting write-up. Assist, Meeting Notes and Chat choose independently.
Built in, no setup
Gemma 4 runs on your Mac and covers 140+ languages. Nothing to configure and nothing to send anywhere.
Or your own
Six local servers (Ollama, LM Studio, llama.cpp server, Jan, vLLM, LocalAI) and seven cloud providers (OpenAI, OpenRouter, Groq, Mistral, DeepSeek, Together AI, xAI) are preset. Keys live in the macOS Keychain.
Claude Code, read-only
Point it at a project folder and the assistant can talk about your actual code. There's deliberately no "allow edits": it's driven by what it hears on a call, which is untrusted input.
Connector tools, one at a time
MCP connectors are denied by default. A Jira connector exposes "create ticket" next to "search issues", so nothing is reachable until you allow it individually.
Sessions, keys and the panel itself
Everything the assistant produces stays on your Mac, and everything about the panel is adjustable.
Searchable session list
Your conversations for the current mode, searchable, renameable and titled automatically. Reopening a saved one never stops a recording that's still running.
Invisible to screen sharing
Excluded from screen capture by default, and now a one-click toggle in the panel rather than a trip to Settings.
Sixteen editable panel keys
Grouped as Actions, Navigation and Display, with Restore Panel Defaults - so nothing clashes with your meeting app.
Compact mode
A smaller panel for a crowded screen, adjustable opacity, and answer text you can size up or down.
What it doesn't do.
Two channels means two labels - Dictera doesn't tell individual voices apart, so named attribution comes from the participant list you type plus whatever the meeting says out loud. Notes are written when you press Stop, not live. Each listening session runs in one language you choose rather than detecting it. No audio is ever stored, so timestamps point at a moment, not a recording. And nothing joins your calendar - you press Start.
Included with every plan.
The AI Assistant isn't an add-on - it comes with Dictera, and the 14-day trial has no limits on it.