These docs follow main: they describe changes no release has yet. Read this page in the latest release

Models and usage

quanthea brings no model of its own. An admin gives it one or more providers in Settings → Model, and Settings → Usage shows what they spent. Pinned dashboards, snapshots, alerts and report runs never call a model.

Both screens are for admins.

Providers

Settings → Model: two providers, one with an API key and one through Entra ID, the provider being edited, and the limits Settings → Model: two providers, one with an API key and one through Entra ID, the provider being edited, and the limits

Add a provider and pick its vendor:

  • Anthropic, OpenAI or Mistral: their own APIs.
  • OpenAI-compatible: any gateway that speaks the OpenAI API, such as LiteLLM, Ollama, OpenRouter, vLLM, or Gemini’s OpenAI endpoint. Give its base URL, or pick one of the common ones. The Gemini one also fills in Gemini’s models for each job.

Then choose how quanthea connects, under Connect with:

  • API key: paste the key. quanthea seals it with its secret key and only ever shows it masked.
  • Identity provider: for a company that bans long-lived keys and puts its models behind a gateway that takes its own sign-in tokens, such as Azure OpenAI with Entra ID, or LiteLLM or Kong behind Okta or Keycloak. See below.

Give each provider a name people recognise (“Mistral free”, “Claude for builds”): threads and usage show it. The providers list says how each one connects, and which still needs a key or secret.

Connect through an identity provider

Settings → Model: a provider reached through Entra ID, with its issuer, client, scope, and where the token goes Settings → Model: a provider reached through Entra ID, with its issuer, client, scope, and where the token goes

Register quanthea as an app (a client) in your identity provider and give it access to the gateway. Then, on the provider:

  1. Set Base URL to your gateway. The vendors’ own APIs take keys only.
  2. Choose Identity provider. Fill in for Entra ID, Okta or Keycloak writes the issuer’s shape; replace the words in capitals with your own.
  3. Give the Client ID and the Client secret, and the Scope or Audience your gateway expects. For Azure OpenAI the scope is https://cognitiveservices.azure.com/.default.
  4. Save, then Test.

quanthea finds the token URL from the issuer’s discovery document. For an issuer without one, choose Give a token URL instead. It asks for a token, keeps it in memory only, and renews it before it expires. The client secret is sealed like a key and goes to the identity provider only.

The test checks the token first, then the model, and says which step failed: a refused client secret, or a gateway that refuses the token, usually because of the scope or audience.

With several providers, one is the default. Editors pick another when they start a thread, and a thread keeps its provider. Editors see the providers’ names and models, never their keys.

The model for each job

Each job can run on its own model. An empty one uses the build model.

JobWhen it runs
Talk the question through and planThe conversation, questions and plans
Build and edit dashboardsWriting panels, alerts and reports
Repair a failed queryThe steps after a write failed, within a run
Titles, tags and descriptionsTagging a dashboard when it is pinned
Answer questions about a dashboardAsk about this, and explaining panels

Building needs a strong tool-calling model. Talking, planning, tagging and answering can use a cheaper one, and each vendor starts them on its cheaper model. The model fields list the vendor’s current models, then whatever the provider’s own API offers; you can also type a model id.

Limits

SettingDefaultWhat it does
Ask to continue after (tool calls)No limitA turn pauses after this many tool calls and asks whether to go on.
Ask to continue after (tokens)No limitA turn pauses after spending this many tokens and asks the same.
Repair attempts per answer3Failed writes the agent may repair before it stops and says why. With 0, it says why after the first one.

A paused turn ends with a card that says what it spent, with Continue. Continue picks up where it stopped, with fresh counts. The limits are for builds left running unattended: leave them empty and a turn runs until it is done.

A stopped build keeps its thread; a reply or Try again starts a new run with fresh attempts. A question about a dashboard has no Continue: past a limit, the model answers from what it has read.

When a provider answers that its quota is spent for the day, quanthea does not retry: it tells the person to try later or to pick another model. A per-minute limit is retried twice.

Behaviour

  • Ask for plan approval before building: on by default. Small edits to an existing panel skip the plan.
  • Test-run every query before showing a panel: failed queries go back to the model with the error, so a broken panel is never saved.
  • Keep the model’s reasoning short: faster and cheaper. Turn it off if a gateway rejects the reasoning effort.
  • Show the queries in a plan: a plan for an existing dashboard also shows each panel’s new query, and how it changes. Plans take longer and cost more.

Usage

Settings → Usage: tokens and list-price cost over 30 days, by feature, with the views of pinned dashboards Settings → Usage: tokens and list-price cost over 30 days, by feature, with the views of pinned dashboards

Settings → Usage shows what the models spent over the last 7, 30 or 90 days, at list prices, and how often pinned dashboards and snapshots were opened.

  • By feature: building dashboards (tagging at pin time included), alerts, reports, questions about dashboards, and panel explanations.
  • By model: each provider and model, with fresh input, cached input, output and cost.
  • By person: the model steps of each person’s threads, questions and explanations.
  • Views: pinned dashboards and snapshots opened per day. No model runs for those.

Usage is kept when threads are deleted and purged, so the history stays whole. Each thread also shows its own tokens and cost: under each answer, and in total in its header.

A model without a known list price is named instead of priced.