Questions

ChimeIn is a shared document that a group builds by talking around one device. Pi tells the voices apart and writes down what matters as you go: an outing plan with real places, a product spec, a chore schedule, a big-purchase comparison, or whatever the group is working out. People heart options and approve on screen or from their phones.

No. Open the link in a browser. There’s no sign-in.

Put the phone or laptop on the table and tap the mic on the home page. That starts a chat, turns on the mic and makes this device the table screen; just start talking and Pi works out what kind of document you need. You can also type what you’re working out instead, or pick a kind first (Plan an outing, a collage, a grocery list, or more under More). Inside a chat, the round + in the middle of the tab bar (or New chat in the sidebar) opens a new one on that device straight away. How it works in the ⋯ menu shows the three steps: plan out loud, everyone scans the QR code, then shape it and decide.

Yes. Tap See demos at the top of the home page and pick a scenario. Lines stream in as if spoken, Pi reads them exactly as it would live, and you watch the document fill in. Pause or stop any time.

On a phone, a chat has tabs along the bottom: Transcript (everything said, and the message box), Live plan (the document, which scrolls to whatever Pi is changing; tap Following Pi to stop and look around yourself, and Follow Pi to go back) and Settings, with a round + in the middle for a new chat. On the table screen, Listen to the room starts and stops the mic from either tab. On a laptop or tablet, your chats and the transcript sit side by side with the live plan next to them. When Pi was asked to find or draft something and has nowhere to look (“find dinner” with no area), the plan waits under a blur with the question until someone answers; any other question waits in the chat while you carry on. You can tick, heart, approve, claim and answer right on the plan; on the table screen, the name chip says whose taps they count as. Add ?display=tv to a chat’s link for a big-screen version for a TV.

Yes. The message box works on the table screen (choose who you’re typing as) and on phones.

Yes. The transcript shows the latest messages first and loads older ones as you scroll up, all the way back to the first line.

The outing planner is the same planner with the same Pi replies, with who’s who coming from voices and devices instead of Apple accounts. The other kinds of documents are new here.

Not yet. ChimeIn is a prototype for trying out the idea with friends and coworkers: a group talks, and Pi builds the shared document. Before it can go to external users it needs real accounts, private sharing, privacy and security work, and native apps. The gaps are listed below.

A privacy policy and terms, a review of what the AI providers see, deleting your own data, moderation and abuse reporting, spending limits per account, servers that stay up and can grow, and the design and accessibility review.

Ready-made ones for a product spec, house chores and a big purchase, plus the outing planner. With “Just start talking”, Pi designs one to fit: a packing list, a hiring scorecard, a potluck sign-up, a cost split, a mood board, a story, a song… The Demo page has examples of each.

Yes. A mood board collects pictures Pi finds on the web (each credited to its page) and photos you add from your phone; a palette keeps the colors. For writing, Pi keeps the story, lyrics or vows in parts and writes or rewrites a part when you ask (“Pi, write the opening”, “give us two endings”); people heart the version they like. Pi doesn’t make images yet.

Yes. It follows your device by default; to pick light or dark yourself, use the sun and moon on the home page or at the bottom of the chats sidebar. The choice is kept on that device.

Yes. Everyone adds their photos from their phone and hearts their favorites; then say “Pi, make a collage of our favorites with the bridge as the big one”. Pi lays them out as a mosaic, grid, film strip or polaroids, adds a title if you name one, and Download saves it as one picture. Drag a photo onto another to swap them.

Pi listens for a few lines, then sets up a document from a library of sections: text, checklists, requirements (must/should/nice), open questions, milestones, tables, options to heart, a weekly schedule, sign-up slots, shared costs, a find-a-time poll, a budget or fundraising goal, a mood board, a photo collage, colors, writing in parts, versions to compare, and a decision to approve. When none of those fit, Pi describes a section of its own (its fields, how it’s shown, and what it adds up), like a garage-sale board by status with money raised. If it sounds like an outing to restaurants or venues, Pi switches to the planner instead. Say “Pi, add a section for …” to change the shape.

No. Talk normally. When something belongs in the document (“it has to work offline”, “Leslie can’t do dishes”, “we can spend about 45k”), Pi adds it quietly and ticks the message. It replies out loud when you ask it to do something or say its name.

Yes. A long line like “I’m coming from Palo Alto, can you find three… uh, four hay mazes near there?” is split into each thing said, with ums and self-corrections cleaned up (four, not three), and Pi handles each one in order. A line that trails off (“near Half Moon…”) waits a moment for the rest.

Cards that still need something are marked: Needs a pick, Tied, 2 open questions or Needs a time. The section chips under the title show a count, and a line like “3 to decide: …” under the title takes you to each one. Moving on to another topic never waits on these; they stay marked until someone decides.

It says so in the chat, instead of staying quiet, whenever the line may have been meant for it.

Each time the document changes, a bubble appears on the part that changed, saying what changed and who it came from.

Yes. Ask for options (“look up some three-row SUVs under 48k”) and Pi fills a table with facts from the web, each with a link to its source; products someone just names get a search link. Ask for real places (“where can I test drive one?”) and Pi adds businesses from Google Maps with addresses, directions links and how far each is from where you said you are. Pi never makes up prices or specs.

Yes. On a day plan, each stop shows the way to the next one: the walk when it’s about 20 minutes or less, otherwise the drive. Tap it to switch to walking, transit or driving (handy if you feel like a long walk). Each place you’re choosing between shows 🚶 walking, 🚇 transit and 🚗 driving times from where you start: tap From ▾ to use where you are now, your home base, someone else’s starting point, or the stop before. In Korea, Google has only transit directions, so walking and driving are estimates marked “about”, and KakaoMap opens for directions.

In Confirmations, at the top of the plan. Pi adds the section as soon as someone mentions booking or buying something (movie tickets, a dinner reservation, a flight, a hotel, a rental) or wants to keep a picture of one for later. Drop in the screenshot or PDF from your phone, or paste the link: Pi reads what it’s for, when, the confirmation number and the seats, and marks that stop on the plan as booked. The picture itself is kept, so a QR code can be scanned at the door: tap Show ticket for it full screen. Anyone with the chat’s link can see confirmations, so leave out card numbers and passwords.

Where you usually start from, like Palo Alto. Set it in Settings (in the sidebar), or tell Pi “my home base is Palo Alto”. For one outing, say “I’m coming from the Castro” instead, or “Mel’s coming from Oakland”. When a plan has real places and Pi doesn’t know where someone is starting, it asks once. Chats see only the name you gave, never the exact spot, and Pi won’t look up a home address.

Settings shows whether this browser can use your location (on, off or blocked), with a button to allow it and how to unblock it. Pi uses it only when you pick “Where I am now”.

Nobody has said where they’re starting from. Say “we’re in Palo Alto” and ask again, and each place shows roughly how far it is (a straight-line distance, not driving time).

Yes. Say who paid for what (“I put the cabin on my card, it was 1,800”), who it was for (“gas is only for the people in my car”) and any uneven shares (“I only came Saturday night”). Pi keeps the costs and works out the fewest payments that settle everyone up; ask “who owes who?”. Tap Mark paid when you pay someone back.

Yes. In the cost split (“Who paid”), tap Add a receipt and take a photo or add the PDF. Pi reads the store, every line and the total, and counts it as paid by whoever added it. Then say who’s getting what: “Jan’s getting the strawberries and kiwis, the rest is mine.” Tax and fees are shared in proportion. Ask “how much should Jan pay me?” and Pi answers to the cent, with the lines it’s for. Whoever pays taps Mark paid once they have.

Yes. “I’ll bring the salad” or “put me down for the morning shift” claims a slot, and “we need 6 people per shift” sets how many it needs. Open spots are flagged. Tap I’ll take it or Leave on your phone.

Yes. Say when you can and can’t (“weekday evenings except Tuesdays”), and Pi records your answer for each option. Everyone answers for themselves. Ask “what works for everyone?” and Pi names the best time; tap your own cells to change your answers.

Yes. Say the numbers as they come in (“I did 12,000 steps today”, “we’re at 1.6M ARR”) and Pi keeps a tracker with a leaderboard, streaks and progress toward the target. Ask “who’s winning?” or “how are we doing?”.

Yes. A roster has a row per person with an RSVP, how much they’ve paid against what’s due, and yes/no checks like “signed the card”. Say “Dev’s in” or “Ruth paid 15”, or tap to update; ask “who hasn’t paid?” or “how many are coming?”.

Yes, in a sign-up. When someone says “I’m allergic to tree nuts” or “I’m vegetarian”, Pi records it, tags each dish with what’s in it, and flags any dish that conflicts (“⚠ nuts: Sam”).

Yes. Tables total their number and money columns and can be held to a limit (“we have about 20 points this sprint”, shown in red when over; ask “are we over?”). Budget lines can be per person and multiply by the headcount (“catering is $110 a head”).

Yes. A budget can be a spending cap or a goal to raise. It shows the running total, what’s left or still to raise, and can hide who gave what.

No. A document has no sign-off: hearts show what the group is leaning toward, and the sections show where things stand. Only an outing’s day plan asks each person to approve a day (“It’s a plan!”). Roles still work: say “I’m the PM” to take one.

Tell Pi the chores, how often and roughly how long, plus any rules (“Sam is away on weekends”, “Leslie can’t do dishes”). Ask for a fair schedule and Pi balances the minutes per person while keeping the rules. Re-balance any time.

Yes. Export at the top of a document gives Markdown (paste into Docs), a spreadsheet (CSV) and a calendar file (dated items and the weekly schedule).

One document per chat. It can’t read people’s calendars (the find-a-time poll is from what you say), it doesn’t make images or music yet, and files other than photos can’t be added. These are planned.

“Find lunch near the game”, “we should go to the Valkyries game”, “push dinner to 8”, “I’m vegetarian”, “what’s the plan”, “undo”, “make it Sunday”, “skip the brunch”, or “plan it for us” for a full draft.

It doesn’t guess your city. Name a place once (“we’re in Palo Alto”) and it uses that.

Yes, from Google Maps, with ratings, prices and walk times. Event dates and venues come from a web search. When Pi isn’t sure, it says it’s a guess or asks.

Yes. Say it as your own pick (“I like the Palisade”, “my vote’s for the second one”, “the beach photo for me”) and Pi hearts it for you; “actually, not that one” or “I changed my mind” takes it back. It only ever hearts for the person speaking, never for someone mentioned (“Ana would love it”), and not for musing (“that one’s kind of cute”). If Pi can’t tell whose voice it was yet, it asks: tap “Heart it as you?” in the chat. Approving is separate: see below.

On a day plan, once every stop has a pick, each person says “Lock it in”. It counts only for the person speaking; “Unlock” takes your approval back. Once Pi has proposed locking in a pick (after a secret vote, or when two or more people back an option and someone names it), “book it”, “do it” or “lock it in” confirms that lock and its calendar invite for the whole group.

Yes. Say “let’s vote privately”, “secret ballot” or “pass the phone around”. Each person picks on their own phone (or taps on the table phone, which hides the choice right away). The group sees only the result, never who picked what; with very few people Pi shows less, so no one’s vote can be worked out.

Pi proposes what follows from it, like “Lock in Deetjen’s for Stay, Friday 4 pm” and a calendar invite, and waits for someone to say “book it” or “lock it in”. Once confirmed, each step shows on the table screen with Undo in case it was too soon. When everything is decided, the table screen says “It’s a plan”.

The place with the most hearts wins. When every stop has a pick, everyone approves and the plan locks (“It’s a plan!”). If it changes later, it asks for approval again.

No. It only plans. “I got seats” just records that someone has tickets.

Yes. A locked day shows Add to calendar: choose Google Calendar to open your own calendar with the outing filled in (every stop, time and address in the notes; tap Save), or the .ics file for Apple Calendar and Outlook. Calendar sections add each event to Google with + Google.

Yes. Say “put these on a map” or “show me photos” and Pi adds a Map section or a Photos board of the plan’s places, from what it already looked up.

Every event and picked place shows its address with Directions (Google Maps, which opens the Maps app on phones) and, on Apple devices, Apple Maps.

A voice model separates up to 8 speakers. Say “Hi, I’m Maya” to name your voice, or tap a voice chip and pick a name.

Similar voices can merge. Pi says so when it notices. On the table screen, tap the name on any message to fix who said it. Phones are the reliable way to vote.

Only within a listening session. If the mic connection drops, it reconnects within 3 minutes and keeps names. After a longer break, say hi again or re-tap your voice.

When you tap Stop (or say “Pi, that’s it”), Pi posts a wrap-up: what you decided, what’s still open and who was there, with a Share with the group button. On a phone it opens your share sheet so you can drop the link straight into your group chat; on a computer it copies the link with a short message. The Share button at the top of the chat and in the ⋯ menu do the same any time.

The link preview shows the chat’s title, where it stands and who’s in it. Opening it asks “Which one is you?”: if you were there, pick your voice and everything you said, hearted or locked in counts as yours; if you weren’t, tap “I wasn’t there: add me”. From then on the chat is in your own sidebar.

Scan the QR code from Share with the group (the ⋯ menu on the table screen) and pick yourself on “Which one is you?”. You can heart, approve, tick items off, and hold the mic button to talk.

Phones need the secure (https) site for the microphone. The live site has it; a laptop’s local Wi-Fi address doesn’t.

Listening stops by itself after 40 minutes, or after 5 minutes with nothing said, so a mic left on doesn’t keep running; tap Listen to the room to carry on. Each chat also has a daily limit on what Pi can spend, and so does ChimeIn as a whole. When one is reached, Pi says so once and stops reading along until the next day (UTC). Hearts, votes and approvals by tap keep working.

On purpose: small talk stays out of the document. Saying “Pi, …” makes it treat what follows as a request. Pi also listens for how a line is said, not just the words: “Pi, can you find dinner?” is a request even when the transcript hears “Hi, can you…”, and “I made an apple pie” isn’t one.

An audio model from Inflection, Pi-D, listens to the last few seconds of each spoken line along with the recent conversation and scores how likely it was said to Pi. A clear yes makes the line a request; a clear no means Pi stays out, even if “pie” was in it. When Pi-D isn’t sure, or doesn’t answer within 1.5 seconds (it runs on a test server that’s sometimes down), Pi falls back to listening for its name. Either way, Pi still writes down what matters from conversation between people.

While listening, the bar above the chat shows Gate: Pi-D (green) when Pi-D is deciding who lines are for, or Gate: fallback (amber) when it can’t be reached and Pi is listening for its name instead. Each spoken line carries a tag: Pi-D 0.97 is Pi-D’s score (green when it counted as said to Pi), and fallback means the name rules decided that line. Hover a tag or the chip for the reason. Demos are scripted text with no audio, so they don’t use Pi-D and show neither.

Pi only takes things out when the words clearly say to (“take the Honda off”, “cross that off”, “rule it out”, “just keep the first two”). If a change would remove something without that, Pi makes the rest of the change and asks, like “Remove Honda Pilot from Options? Say “yes” and I’ll take it out.” Anyone’s yes within 5 minutes removes it; a no keeps it.

Say or type “undo”. It takes back your own last change, unless someone changed that part after you.

Each chat is separate, and everyone with its link shares it live. There’s no public list; “Recent” on the home page shows only the ones your browser started or joined.

Anyone with the link can open it and add to it, so share it with the people it’s for. Links aren’t listed anywhere or shown to search engines. There are no passwords or view-only links yet.

No. Audio is transcribed live and only the words are kept. To tell whether a line was said to Pi, the last 8 seconds of room audio before it are sent to Inflection’s Pi-D service with the recent lines; ChimeIn doesn’t store that audio.

As files on our Google Cloud server, one file per chat. Photos people add are stored with that chat, can be seen by anyone with its link, and are deleted with it. The server’s log (which includes short snippets of what was said, to trace problems) is kept for 14 days.

If you started the chat: ⋯ → Delete for everyone. It’s permanent and immediate, and everyone you shared it with loses it too. Anyone else sees Remove from my list, which only takes it off their own device. Only the device that started a chat can delete it (chats started before this change can still be deleted by anyone with the link).

In the outing planner, your personal budget is hidden from everyone else (without sign-in, someone could still pick your name on a phone). A document’s Budget section is shared with the group.

The document and what was just said, plus each photo someone adds (Claude writes a short caption so Pi can find and group it). For documents that includes people’s names next to their words, so Pi can credit who said what and record roles and owners. The planner’s AI draft sees recent talk labelled “A:”, “B:”, never names. Inflection’s Pi-D sees the last 8 seconds of audio before each spoken line, that line, the six before it with speakers’ names, and what Pi last said.

About 4–6 people talking into one mic, 8 at most. Many more can join and vote from phones.

A few live mics at a time, a limit of the shared voice service. When every spot is taken, the next group waits in line: the bar shows “In line for a mic” with your place, and listening starts by itself when a spot opens. You can type in the meantime.

The mic reconnects on its own. Everything is saved on every change.

Messages are read one at a time per chat so they stay in order. Each spoken line first waits for Pi-D to say whether it was for Pi (usually under a fifth of a second, never more than 1.5). Web and Maps lookups take a few seconds, sometimes 20 or more.

About $15–20 a month for the server, plus usage: AI calls (a few cents per conversation), Google Maps (capped at $2 per chat per day and $10 per day overall), and web lookups (under a cent each).

Inflection’s diarization and speech recognition for voice; Inflection’s Pi-D (an audio model on a test server) to tell when a spoken line is said to Pi; Claude to design documents, turn talk into edits and draft plans, and a faster Claude model to split long spoken lines into each thing said. The quick “is this worth writing down?” check runs on Nemotron when the room stack is connected, otherwise on Claude. Ask about ChimeIn at the top of this page for which model does what on this server right now.

Not yet. Its replies appear in the chat. Speaking aloud needs a voice-cloning text-to-speech service from the voice team.

Not yet. With no sign-in and no password, anyone with the URL can use the app’s AI and Maps keys. Before wider sharing, it needs sign-in or a usage cap.