Quorum runs as a hosted MCP server. Claude for Excel can add it as a custom connector, which means a panel of frontier models becomes something you can convene from inside a workbook, on the questions a spreadsheet cannot answer for you.
This is a connection we have built and verified from our side, not a pattern with users behind it. There are no customers running Quorum in Excel today, no case study, and no adoption number we could honestly quote. If you wire this up, you may well be the first person to do it. The page is written that way on purpose. Tell us what breaks.
Two reasons, and only the second one is interesting.
The first is that the work already lives in the sheet. The forecast, the pricing model, the supplier comparison, the headcount plan. Moving that into a chat window to ask a question about it is friction, and friction is why good questions do not get asked.
The second is the shape of the questions. A spreadsheet is very good at lookups and arithmetic, and it has been good at them for forty years. It cannot tell you whether the assumption in row 4 is defensible, or which of three options to take when the numbers land within noise of each other, or what you have failed to model at all. Those are judgement calls. They are the questions people actually stall on, and they are exactly the class of question where several models arguing beats one model answering confidently.
That is the case for putting a deliberation next to a workbook. Not to compute anything. To disagree with you before a number becomes a decision.
Quorum is a remote MCP server. There is nothing to install, no package, nothing running on your machine.
Server URL
https://www.quorum.dog/mcpAuthentication is a static bearer token: your Quorum API key, sent as a header.
Authorization: Bearer qk_...Claude for Excel accepts a static bearer token. It does not require OAuth, so there is no app registration step, no redirect URL and no consent screen to configure. Add Quorum as a custom connector, give it that URL and that key, and the tools appear.
We are deliberately not printing click-by-click instructions here. Connector UIs move, and a stale set of menu names is worse than none. The outcome you are aiming for is simple: a custom connector pointed at the URL above, carrying your key in the Authorization header.
To check the key itself before involving Excel at all, list the tools directly:
curl -sS -X POST https://www.quorum.dog/mcp \
-H "Content-Type: application/json" \
-H "Authorization: Bearer qk_..." \
-d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'The same key works against the REST API, which is OpenAI-compatible with a base URL of https://www.quorum.dog/v1. See the API reference.
Four of them are free, and two spend money, which is most of what you need to know.
| Tool | Cost | What it is for in a workbook |
|---|---|---|
estimate | Free | Classify and price a question without running it — and get deliberation_value, a read on whether a panel is likely to help at all. Ask it first when you are not sure the question is worth one. billed_usd comes back 0 on every call, by design. |
list_modes | Free | Which panels your key is allowed to convene. Modes differ in who sits on the panel and how hard they argue, so this is the menu. |
deliberate | Real money | Convene the panel on the question you paste in. This is the one that costs, and the one you wait for. |
get_receipt | Free | For a call you already made: which model sat in each seat, the judge scores, the cost and the latency. Use it when you want to know whether the panel actually disagreed or just nodded along. |
run_test | 25% off deliberate, per question | Run your own question buckets, judges and repeats against a Mode. Rarely what you want from a spreadsheet, but it is there. dry_run: true prices a run without spending anything. |
certify_mode | $50 flat | A fixed 150-question certification run against a Mode. It takes about two and a half hours, so it is not something to trigger from a cell by accident. |
run_test and certify_mode are not read-only. certify_mode commits $50 the moment it launches, and run_test bills per question. Both were added to the hosted server on 2026-08-22, so an assistant connected to Quorum can reach them from Excel. Neither is something to call speculatively while exploring what a connector does.
Say you have an eighteen-month revenue model. Growth is compounding at a rate you typed in during a meeting three weeks ago and have not revisited. Everything downstream depends on it. Nobody has argued with it.
The pattern is: summarise the assumption and its context in a paragraph, then ask the panel to attack it.
Estimate first, then deliberate.
estimate:
"We are modelling 8% month-on-month new-revenue growth for
18 months. B2B SaaS, average contract £14k, two-person sales
team, no paid acquisition, 62 customers today. Is that
assumption defensible, and what would have to be true for
it to hold at month 12?"
-> difficulty_score high, task_type strategy, a price you can see
before you spend it.
deliberate:
same question, same paragraph.
-> the panel's answer, plus a request_id.
get_receipt(request_id):
-> who sat on the panel, how much they disagreed, what it cost.What comes back is not a number to paste into a cell. It is the argument you would have had if you had three sceptical colleagues to hand. In our own use the useful part is rarely the verdict. It is the objection nobody in the room had thought of, surfaced because one seat on the panel would not let it go.
The other honest use is the tie. Three suppliers, three vendors, two hiring plans, and the spreadsheet has done its job and told you they are within a few per cent of each other. The numbers have stopped deciding. Paste the comparison and the constraints in, and ask the panel which one it would take and what it is trading away. A single model will pick one and sound certain. A panel will usually tell you the choice is closer than you want it to be, and which risk you are actually accepting.
If you want the deliberation to happen automatically only on hard questions rather than on every question you type, that logic belongs on your side. The escalation router guide covers the shape of it.
estimate exists so that the price is never a surprise. It classifies the question, prices what the real call would cost, and does not run it.
| Measured on production, 2026-08-20 | Latency | Cost |
|---|---|---|
estimate | ~0.9 s | ~$0.00006 |
deliberate, median (p50), n=3,845 | 28.5 s | Varies by mode; estimate tells you |
deliberate, p90, n=3,845 | 88.2 s | Varies by mode; estimate tells you |
Read the cost column carefully. The $0.00006 is what it costs us to run the classification, not what you pay. You pay nothing for estimate: billed_usd is 0 on every call. It is free because it is nearly free to provide, which is a more durable reason than generosity.
So the discipline is cheap. Classify for a fraction of a penny and a second of waiting. Look at what it says the question is and what the answer would cost. Then decide whether to spend it. Full pricing is on Pricing, explained.
This is the section that decides whether the rest of the page was worth reading.
It sees the paragraph you send it. Nothing else. It cannot read cells, walk formulas, follow a reference to another sheet, or inspect a range you have not pasted. There is no live connection to the file. If the assistant in Excel gathers context from the sheet before calling Quorum, then what Quorum receives is that assistant's summary, and a summary is where detail goes to die.
Which makes the round trip lossy at both ends. Summarised question in, summarised answer out. That is a real cost, and it means the pattern suits questions about judgement far better than questions about data. If the answer depends on a precise figure, put the precise figure in the paragraph yourself rather than trusting it to survive the trip.
28.5 seconds at the median, 88.2 at p90. That is the honest number and it is not going to feel fast inside a spreadsheet, where most things happen instantly. Half a minute of nothing is a long time when you are used to F9. Treat a deliberation as something you ask for deliberately, then go and do something else.
Convening several frontier models to confirm a VAT rate or convert a currency is a waste of your money and a minute of your life. Ask that of anything else. Quorum earns its cost on the questions that have no lookup answer, and it is better for both of us if you use it that way.
Worth repeating at the end. The connection works and the numbers above are measured. Whether a panel in a spreadsheet is genuinely useful day to day is an open question, and we do not yet have a single external user's answer to it.
Connect Quorum to your editor for the full MCP contract and other clients · The escalation router for deciding when a panel is worth it · Pricing, explained · API reference · Help.
This page gets Quorum reachable from a spreadsheet. AI Deliberation & Excel is the other half: which analyst jobs are worth a 28-second panel, which are not, and what the round trip costs in fidelity, with the worked exhibits behind each. 13 pages, PDF.
Every paper in the series sits on the whitepaper shelf — one per surface.