Ellipses 2.0 listens to your live call through Otter and puts the right thing in front of you at the right moment: the correction, the number, the better phrasing, the question worth asking. Grounded in your own email, Slack, past meetings and store data.
Runs in your own Claude account, on your own connectors. Installing takes a few minutes.
Ellipses 2.0 reads the live transcript every 20 to 45 seconds, compares what was just said against a short briefing you control, and returns at most three cards per pass. Each card is color-coded by what it needs from you, carries the words you could say next, and shows exactly where its substance came from. It assists you; it does not hide from the room.
Ellipses runs inside Claude, on your account and your connectors. A shared link can't lend you someone else's, so you install your own copy. It takes a few minutes.
Publish the attached ellipses-2-app.html as a new artifact titled "Ellipses 2.0". Publish the file exactly as it is. Do not rewrite, restyle or shorten it. Use the latest runtime contract and declare these capabilities:
{"sample":{},"room":{},"downloads":true,"user":{"scopes":["profile"]},"db":{"rules":[{"path":"","read":"view","write":"admin"},{"path":"data/users/{self}","write":"interact"},{"path":"stats","read":"admin","write":"admin"},{"path":"stats/{self}","read":"interact","write":"interact"}]},"mcp":{"servers":[{"server":"Otter.ai","tools":["otter_search","otter_fetch"]},{"server":"Gmail","tools":["search_threads","get_message"]},{"server":"Slack","tools":["slack_search_public_and_private"]},{"server":"Google Calendar","tools":["list_events"]},{"server":"Google Drive","tools":["search_files"]},{"server":"Parallel Search","tools":["web_search"]},{"server":"Shopify","tools":["run-analytics-query"]},{"server":"Amazon Selling Partner","tools":["accounts_getMerchants","call_read_only_tool"]},{"server":"Stay","tools":["query_analytics"]},{"server":"Intuit QuickBooks","tools":["profit_loss_quickbooks_account_text"]}]}}
Declare every connector in that list, including any I have not connected yet. The app marks a missing connector as unavailable. When it is published, give me the link and tell me whether every capability was accepted.
Every connector tool in the prompt is read-only: search and read email, search Slack, read Otter transcripts, list calendar events, search Drive, search the web, run store and finance reports. The app can't send, post, edit or delete anything.
| Tier | When it appears | What to do |
|---|---|---|
| Urgent | Wrong information stated as fact that affects a decision | Say something now |
| Check | Unverified, vague or unsourced claim | Confirm it, or run a deep check |
| Answer | Someone just asked a question the briefing or data can answer | Read it, say it |
| Add | Helpful context you may want to bring in | Use if the moment allows |
| Coach | Listening, framing, de-escalation, persuasion | Adjust how you're showing up |
| Loop | A commitment made, or a decision with no owner or date | Close it before the call ends |
Six tiers, six colors, one question each. Urgent cards pin to the top of the stack. Any tier can be muted for the rest of the call, and any card can be pinned to a section you return to later.
Every card carries two to four chips matched to the meeting type. Work calls: Check with data, ELI5, Devil's advocate, Suggested response, Add to follow-up. Personal calls: Look it up, Good follow-up question, Remember this.
Pick how hard Ellipses works. Manual when you'd rather tap. Standard for most calls. Turbo when the stakes are high and the room moves fast. Auto mode pauses on its own after a set time, after five minutes of silence, or ten minutes past the calendar event's end, so a forgotten tab never runs up your usage.
Otter's transcript lags speech by up to about ten seconds, so polling faster than 20 seconds adds cost, not speed. Use SOS or the paste box when you need this exact moment.
No automatic pulls. Tap Check now whenever you want a pass.
Balanced triage on each pass. The right default for most calls.
Quick-tier triage. An Urgent card with an incorrect verdict auto-escalates to a deep check, capped at three per ten minutes.
Sets the model tier for deep checks and answers. If your plan substitutes a tier, the card says which one answered.
For when you're called on and lost. One tap pulls the transcript through to now, then returns a natural bridge sentence that buys a moment, followed by a short answer to what was most likely asked. Copy Meet's captions into the paste box, press Enter, and it runs on that text directly.
One button for the whole call: three to five options the room hasn't considered, each with a one-line trade-off. Keep any of them into your notes.
Ask anything mid-call. Fast answers from the briefing and transcript, or turn on connectors for a sourced answer in 30 to 90 seconds.
Every output carries a lightweight badge saying where it came from. Any number without a tool or briefing source gets the unverified badge automatically. There is no confidence score to interpret; the badge is the whole story.
Work calls default to as many sources as the plan's tool cap allows, in priority order: Gmail, Otter, Slack, Amazon, Shopify, Google Drive, web search, Stay, QuickBooks. A visible count means nothing is ever dropped silently. Calendar detection and the Otter transcript never take a tool slot. Personal calls use no connectors at all.
Ellipses runs fine without Amazon data. If you sell on Amazon, connect Amazon's official Selling Partner connector in claude.ai under Customize → Connectors, and deep checks can pull your sales and FBA inventory. There's nothing to paste: Ellipses finds your seller account and marketplace by itself, the US marketplace first.
Amazon's own connector for Claude. Sales, units, orders and average unit price for any date range in the last two years, for the whole store or one ASIN or SKU, plus FBA inventory with inbound and reserved units. Read-only, with no Amazon Ads or traffic data.
Amazon's announcement and setupWhen the page's limits start to bind, v3 rebuilds on the Claude API: the same connectors, built-in web search, long-term memory per user, and direct audio to remove the transcript lag.