Frequently asked questions.

What a call costs depends on the Mode and the question. An API-key call bills your organization the engines at what their provider charges, the platform fee and the Mode's fee. On the web, a question draws on your plan at real cost. POST /v1/estimate is free and quotes the whole call before you spend. The details are on the Pricing page.

No. Quorum does not use your inputs or outputs to train AI models, and does not use your content for cross-user learning of any kind. The Privacy Policy and Terms of Service have the full terms.

It depends on the engines the Mode you are using assigns to its seats. The current engine catalog uses these providers: Amazon Bedrock, OpenAI, Google, Groq, Vercel AI Gateway. The list changes when the catalog does. Grounding uses Serper (web search queries, when grounding runs).

The company that made a model is not always the company that runs it. Seats reached through Vercel AI Gateway are routed by Vercel to a company that serves that model. The models available this way come from these makers: OpenAI, Alibaba, Mistral, MiniMax, Amazon, Inception, NVIDIA, Meta, Google, inclusionAI, Inference.net, Morph, Moonshot AI, ByteDance, Interfaze, Cohere, Mixedbread, Arcee AI and DeepSeek. The engine on each seat is named on the receipt.

Prompts sent with your keys go directly to that provider under your agreement with them, and their terms apply. See Bring your own keys.

Delete a single conversation from the left navigation in Chat. To delete conversations from API, MCP or connected-app calls, or your whole account, email support@quorum.dog. Deleting removes the content from the active database. Backup copies can persist for up to 30 days before they are removed from backup systems.

Your content and account data are deleted within 30 days, except where the law requires us to keep something.

Each deliberation run with an API key, through MCP or through a connected app is saved as a conversation linked to its call record. Free estimates, and calls that fail before a panel answers, leave a call record and no conversation. The conversation holds your question, each seat's answer, the judge's verdict when a judge ran, and the final answer. The call's receipt names the engine on each seat; GET /v1/receipts/{request_id}?include_transcript=true also returns each seat's answer. These conversations do not appear in your Chat history. To have them deleted, email support@quorum.dog.

Each call also has a call record. For API-key calls, the record holds your organization and API key IDs, the Mode and price tier, the idempotency key, status, error code, rounds, seats, token counts, latency, timestamps, and the call's cost, fees and billed amount. It also keeps the answer text when the call stored one, so a retried request with the same idempotency key returns the original answer. For connected-app calls, the record holds your user ID, the Mode, which account paid, status, error code, rounds, seats, token counts, latency, timestamps, cost and billed amount.

We use these records for billing and your usage pages, reliability monitoring, rate limits, usage analytics and replaying retried requests. Account data, such as your email and billing identifiers, is kept for the life of the account. Questions and deletion requests go to support@quorum.dog.

Delete single conversations from the left navigation in Chat. To delete conversations from API, MCP or connected-app calls, delete your account, or request a copy of your data, email support@quorum.dog.

An engine is a model. A source is a provider and endpoint that serves it. One engine can have many sources.

It sets aside unhealthy sources, ranks the rest by what the call costs the caller including the platform fee, and breaks ties on latency.

The call goes to another source for the same engine. If none can run, it goes to a backup the Mode approved. If that fails too, the seat is shown as unavailable on the receipt.

When a Mode is resolved, the engines it approved: each seat's engine and its backups.

A sync scheduled once a week finds new models. Each one with an exact price goes live once a probe passes it. Sources that stop passing drop out of ranking, and a probe job scheduled every 15 minutes re-tests them.

A source serves the same model. Sources differ in price, latency and health.

Add https://www.quorum.dog/mcp to the app, then sign in to Quorum and approve; Grok, Claude Code and GitHub take an API key instead. The steps for each app are in Connect Quorum to your tools.

Grok, Claude Code and GitHub. Issue a key with + Issue API key on Usage and send it as a Bearer token.

Ask the assistant to call list_modes. It is free and returns the Modes you can call.

Use OAuth when the app supports it: paste the endpoint, sign in to Quorum and approve once. Use an API key for Grok, Claude Code and GitHub, for server-side code and for scripts. Setup for each app is in Connect Quorum to your tools.

A connected app on your plan pays real cost, with no Mode fee.

API keys always bill your organization's API account. Issue keys with + Issue API key on Usage. Settings, API keys is the page for your own provider keys.

Change the client's base URL to https://www.quorumllm.com/v1, use a Quorum API key, and set model to a Mode ticker such as QRUM:STAN. The request and response follow the chat completions shape. Set the client timeout to 300 seconds and leave streaming off.

The ticker of a Mode, in the form COMPANY:MODE, for example QRUM:STAN. Call GET /v1/models to list the tickers your key can use. Over MCP, list_modes returns them.

Yes. The official openai packages for Python and JavaScript take a changed base URL. The quorum block arrives as an extra field on the response object: r.model_dump()["quorum"] in Python, r.quorum in JavaScript.

A quorum block with the request id, whether the seats converged, what remained contested, the model ids that ran, and the billed cost. The request id opens a receipt that lists every seat.

The request had stream: true. Remove it or set it to false.

One model answers once, and its confidence is the only signal you get. A reasoning model thinks longer, which makes the same point of view more thorough. Deliberation adds a second point of view from independent seats, and shows you where they agree and where they split. Reasoning suits answers that are hard to compute. Deliberation suits answers that are hard to be sure about. A reasoning model can sit in a seat and runs at the Mode's thinking level, with token limits set by the Mode.

The seats answer independently, and the judge scores them against each other. Open the receipt: it lists each seat's engine and judge score, so you can see which engines sat, and quorum.deliberation lists what the judge read in each round. A Mode's maker chooses the engines and can turn on family diversity, which keeps Auto seats from sharing a model maker.

One pass in which the seats answer. The judge reading that pass is part of the round. Synthesis is a separate step after the rounds.

The seats did not settle on one answer. The question is contested, and the receipt shows where they split. quorum.converged is set on every response that is not a replay. It is null when no judge read the round, as on the express lane.

Convene a panel for judgement calls: a consequential assumption, options close enough that the numbers no longer separate them, a diagnosis you cannot check, a decision you will have to defend, or a "what have I missed" check. Use a single model for lookups, syntax, formatting, conversions, arithmetic and anything you need reproduced word for word. Call estimate first. Its deliberation_value is one of likely_helps, likely_hurts, unknown, and says whether a panel is expected to help on your prompt.

No. A Mode can turn on the express lane, where one seat answers an easy question. In Chat, choosing Foundation or Frontier convenes the panel for that question.

The seats and their backups, depth, the critique policy, the engines that judge and synthesise, the express lane, the thinking level and a price.

Yes. You can run it once it is saved in the Builder, or released if you saved it over MCP or with an API key. Your teammates can call it once Callable by API is on.

It picks an engine for each question and the Mode stores the rule, so the pick follows the ranking as engines change. An Auto seat resolves at call time to the best active engine inside the caller's tier ceiling, ranked by the AA score for the question's task type. With family diversity on, it skips engines from a maker another seat already holds.

It keeps an Auto seat from sharing a model maker with another seat. Opt-in per Mode: off unless the maker ticks it, and only with at least one Auto seat. Engines you picked yourself count as already taken.

Yes. Over MCP, genie_propose_mode drafts a spec, validate_mode checks it, save_mode saves it and list_mode lists it. run_test, ask_genie, approve_quote and certify_mode cover the Lab and certification.

Asking the Genie is free. A Lab or certification run is priced as a quote and launches when you approve that quote.

It stays pinned as the public card and shows its certification date. A new certification replaces it. Your saved edit goes live at once for every caller of the ticker, Marketplace callers included, while the card still describes the certified version. Before you save changes to a certified, listed Mode, test them in the Lab, and certify again once they are saved.

A saved configuration of how Quorum answers: which engines take which seats, how much critique runs, how many rounds, and the Mode's fee. Every answer runs inside one.

Standard seats three engines and critiques on harder questions. Einstein seats frontier engines and critiques on every question.

Read its card in the Modes gallery. It lists the providers, the seat count and the critique level. The maker's note on when to use it is quorum.use_when in list_modes.

Yes. Saving a Mode needs an organization account. Publishing to the Marketplace uses the same organization. Keep the Mode private, share it with your team, or publish it. Start in the Mode Builder; Make a Mode has the steps.

A Mode listed in the Marketplace by Quorum or by an outside Mode Maker. Anyone else can use a Mode once it is listed and certified.

Free $0, Plus $20, Pro $45, Teams $30/seat. Annual billing: Plus $16.67, Pro $37.50, Teams $25 per seat; billed annually at $200, $450 and $300 per seat.

It is the annual price divided by twelve. You pay the annual price once a year.

3 seats. The price per seat is the same at any size.

A percentage of what a seat's source billed for that seat: 5% when the seat runs on Quorum's keys and 0.5% when it runs on your own key. It applies to API calls only.

The fee applies to API calls. A connected app that bills your plan pays real cost.

Use Turbo, or stop until your allowance resets or you top up.

An API-key call has a ceiling: the higher of $2.00 and the price of the Mode being called, unless the API key carries its own ceiling. A request can set a lower ceiling with quorum.max_cost_usd. Calls on your plan have no per-call ceiling.

Yes. POST /v1/estimate is free. It quotes the whole call for your question: engines, platform fee and the Mode's fee. It also carries a verdict on whether a panel is likely to help.

No. Every plan runs on Quorum's keys. Add yours when you want a maker to bill you directly and those seats to come off your plan meter.

OpenAI, Google, xAI. Settings, API keys shows the current list.

No. A seat that answers on your key is not counted against your plan.

On API calls, a platform fee of 0.5% of what the source billed for the seat. The fee applies to API calls only.

Keys are write-only. The list in Settings shows which makers you have keys for. To change a key, add a new one for the same maker.

The seat is skipped. With If my key fails, use Quorum's sources and bill me on, it runs on Quorum's keys instead.

In Chat, select the Quorum mark under the answer. Over the API or MCP, call get_receipt with the request_id from the original call.

The platform fee on that seat: a rate applied to what the source billed. The rate is 5% on Quorum's keys and 0.5% on your own key. It applies to API calls only. Pricing has the full bill.

The Mode assigned an engine to that seat, and neither that engine nor the engines the Mode approved as its backups had a source that could run. The receipt prints the reason.

No fee was recorded for that seat. Calls billed to a plan record none.

The call went over its ceiling. The amount over is credited back in cap_credit_usd, and the receipt status reads capped.

They are a per-question choice of which engines fill the seats. Express uses small, fast engines, Foundation mid-range engines and Frontier frontier engines. Your plan sets the highest one you can use.

Express for everyday work. Foundation when you want the panel to debate the question. Frontier when a wrong answer is expensive.

No. Your plan sets the highest tier you can reach. You pick the tier for each prompt, up to that ceiling.

Yes. Express is available on every account, Pro included.

Chat opens the register or upgrade prompt. A connected app billed to your plan returns tier_not_permitted with the highest tier your plan allows. An API key accepts any listed tier.

Read the receipt. It lists the engines that ran on the answer and whether a deliberation happened.