Repository navigation
Pennant 0.3.0
Pennant 0.3.0
Talk with Pennant out loud, in a voice that runs on your Mac, and a quicker Pennant chat.
- Talk mode. The waveform button in the Pennant chat (⇧⌘T), on the Mac and the iPhone, is a spoken conversation. Nothing you say goes to a cloud service.
- Listening: Apple's speech recognition hears you on the device. A short pause sends what you said (1.5 seconds by default, or 1 to 3), and a breath mid-sentence doesn't.
- Answering: Pennant answers in a few spoken sentences and says a few words before it goes off to work. A long message, such as a report or links, is said in a sentence or two of what happened and what you need to do. Later news from threads is read too.
- Talking over it stops it, and ⇧⌘M mutes you while it keeps talking.
- It works on Macs where echo cancellation won't start, for example with Microsoft Teams' audio devices installed.
- A natural voice that runs on your Mac. On Apple silicon, Talk mode speaks with a neural speech model on the Mac's GPU, in Pennant Voice, a helper inside the app. Download a voice once in Settings › Pennant › Talk mode:
- Kokoro is light and quick, with American and British voices.
- Penny is a woman's voice, cloned from a short recording that ships with the app. A reply goes to her in one piece.
- Expressive is Qwen3-TTS 1.7B: Ryan and Aiden, plus seven speakers with Chinese, Japanese and Korean accents.
- On the iPhone the Mac makes the speech and streams it to the phone, so the phone hears Penny and the Qwen voices with nothing to download.
- Without a natural voice (Intel Macs, or before a download) Talk mode uses the system's voices, Siri's natural ones first.
- For skills,
Pennant Voice speakandtranscribemake speech and turn audio back into words, from the command line.
- A quicker Pennant chat.
- Replies think lightly and carry a short list of tools. Work past a few steps goes to a thread.
- The chat looks things up in Pennant's own Chrome tabs, so search engines don't block it.
- Start over, next to Compact, gives you a fresh conversation. Pennant's memory and the work in threads stay as they are.
- Threads that stay on task. A thread the chat starts gets 50 steps, and you can say "keep going". When something it needs isn't there, such as a connection, an account or access, it stops and says what's missing instead of working around it.
- The heartbeat reads the inbox. With Microsoft 365 connected, each check-in looks at unread mail since the last one and mentions only what needs you soon. A check-in with nothing new still costs nothing.
- Fewer delete cards.
- Cleaning up a temporary folder no longer asks, even when the folder is in a variable or made with
mktemp. - A delete card for a path has Allow deletes here for this thread, which covers later deletes in the same place in that thread.
- Cleaning up a temporary folder no longer asks, even when the folder is in a variable or made with
- Models are called by what they run. A model's name is worked out from the model as its provider names it, where it runs, and its effort, e.g. "DeepSeek-V4.1-Flash · Azure my-ai-resource · medium". A quick double change on the Models page could make the default model run another's settings under its name. It can't now.
- Smaller fixes:
- A reply's first word no longer sits on a line of its own while the reply streams.
- In Chrome you see one Pennant pointer at a time.
- Pennant works in Mac apps in the background instead of giving up when you're using the Mac.
- The sidebar's More opens a panel, and Settings has a row of its own.
pennant agent edit --stylesets Pennant's manner from the command line.
Requires macOS 15 or later, on Apple silicon or Intel; natural voices need Apple silicon and download 340 MB to 4.5 GB each. Talk mode asks for the microphone and speech recognition. This is a preview; expect rough edges, and please report them.