Skip to content

AI settings

AI settings are how the hive runs models: provider, key, tier, usage, and a monthly token cap. They sit under the platform plan, not beside it. There is no “AI plan” and no second invoice from InHive for tokens.

Only admins and managers can change them. Open the avatar menu → Organization settingsAI. Members see a reduced view — no keys, no budgets.

InHive hosts speech-to-text for calls (Whisper). Every other model call — Ask, the assistant, digests, specs, suggestions, meeting actions, automations, knowledge embeddings — goes to your provider. InHive never spends a platform key on the hive’s work.

PlanWhat you need
StudioYour own provider key. Default cap 5M tokens per UTC month (you can raise it).
TeamYour own provider key. Default cap 25M tokens; frontier tier by default.
EnterpriseYour own provider key. Same product as Team; set the cap to match the contract.
TrialFull Team for 14 days. Still paste a hive key before anyone relies on AI.

If a feature is greyed out with an upgrade prompt, that is the platform plan, not a missing key. If Ask or the assistant says AI is not configured, that is a missing key.

  1. Pick an API provider. OpenRouter is the usual choice — one key, many models. Direct vendors (OpenAI, Anthropic, xAI, Google, DeepSeek, Moonshot, Zhipu, Qwen, Mistral, Groq) need a key from that vendor’s console.
  2. Paste the API key and save. Required for every provider. The hive stores it encrypted. The browser never gets it back — only “a key is set.” Leave the field blank to keep the current key. Remove clears the key for the current provider only.
  3. Choose a model tier: Economy (fastest / cheapest), Standard (balanced), or Frontier (highest capability). The plan sets a default; changing the plan will follow that default unless you already picked a tier yourself.
  4. Optionally set models by capability — leave a row on “Tier default,” or pick a catalog model (or a custom provider/model-id) for that kind of work. Groups cover digests, Ask, drafting, and the assistant. Open Per request type only if one capability needs a different model than its group.
  5. Set the monthly token limit (or leave it empty to use the plan default). See Token spending below.
  6. Test connection. A pass means the current key and model answered. A fail is the key, the provider, or the network — not the platform plan.

If you run Claude, Grok, Gemini, or another vendor without an embeddings API, also store an OpenRouter or OpenAI key. Knowledge ingest and search embeddings use that key. The completion provider stays the one you picked.

The embedding model itself is fixed. There is nothing to set for it — changing it would break search against existing knowledge.

AI display name is optional. When set, people see it on Ask and the assistant. When empty, those surfaces are nameless — the label is just Ask. This is a label, not a second product.

Token usage is the UTC calendar month against the hive cap. You pay your provider; the cap only stops the hive when the month is used up. Managers set the cap under Organization settings → AI. Empty uses the plan default (Studio 5M, Team/Enterprise 25M). 0 pauses AI for the rest of the month.

Counted against the cap:

KindWhat burns it
CompletionsAsk, the assistant, digests, specs, Suggest with AI, meeting actions, automations, onboarding proposals
EmbeddingsKnowledge ingest and search (files, notes, chat archives, transcripts)
MCPAgent tools that call the same models — they share this cap. A looping agent will burn it.

Not counted: call speech-to-text (self-hosted Whisper). Pasted/imported meeting text still uses completions if you run AI actions.

Managers get an alert at 80%. When the cap is exhausted, AI calls stop until next month (or you raise the limit). There are no overage packs and no metered AI SKU from InHive.

Past-due billing keeps AI going for a short grace window, then pauses. See Billing.

Two collapsed tiles on the same AI page:

  • Memory health — useful Ask and assistant answers that cited a source, by week, person, and project. Askers tap useful / not useful under each answer; this is that score.
  • Corpus health — what is actually in project memory (files, notes, specs, transcripts, …), plus holes and skipped files. Re-index missing queues sources that never made it in.

If answers are thin, the usual fix is the knowledge base, not a different model.

Scheduled workflows and their trust policy live under Organization settings → Automations, not here. See Automations.

Plans · Knowledge · Assistant · Agents