Logo StartupKit
EN
English Deutsch Français Español Polski

AI Providers and Spend Controls

Choose where Kit sends in-app AI work, manage provider keys and models, and cap account-key spend.

What This Controls

Integrations > AI Providers controls most generative AI work run inside Kit. It is separate from connecting an external assistant through MCP: MCP controls what the assistant may do in Kit, while AI Providers controls which provider and key Kit uses for chat and generation tasks. Compensation Research background enrichment, Hiring Live Review, and semantic indexing and search have separate processor boundaries described below.

Only account admins can change these settings.

Providers and Key Ownership

Kit supports Google AI (Gemini), Anthropic, and OpenRouter. A provider can use either:

  • Kit’s platform key, when one is available. Kit enforces its platform-key credit limits for chat and generation calls; embeddings are an exception described below.
  • Your account key. Requests go against your provider account. Kit encrypts the key at rest and shows only a short hint after saving it.

Saving a key makes that provider available; it does not automatically make it preferred. Kit validates a new key with a small provider request before storing it.

Removing your account key returns that provider to Kit’s platform key when one is available. If no key remains, that provider cannot be selected. Removing the currently preferred provider key also clears that preference when the provider is no longer usable.

Choose the Preferred Provider

The preferred provider is the default for new in-app chat and generation work, except Compensation Research background enrichment and Hiring Live Review. It does not change the embedding provider. Kit uses the preferred provider only while a working key is available. The dashboard shows which providers are configured and which one is preferred.

For OpenRouter, account admins can choose models for the premium, standard, and economy task tiers. Kit validates saved model identifiers against OpenRouter’s available tool-capable text models. If a selected model later disappears, Kit returns that tier to its maintained default and notifies admins.

Spending Caps

Account-key caps are budget fences inside Kit; they do not change the limit at the provider.

  • A monthly cap is required when setting custom limits.
  • A daily cap is optional.
  • The first account key starts with a monthly Kit cap; review it before enabling production workloads.
  • Removing the caps removes Kit’s account-key fence. Provider-side budgets and billing alerts remain your responsibility.

Except for Compensation Research background enrichment and Hiring Live Review, Kit records chat and generation requests, tokens, estimated cost, provider, model, key source, and the initiating member where available. Some account-scoped embeddings are recorded separately as estimated Gemini platform-key usage; see the data boundary below. The AI settings page shows current-month usage by member and monthly usage history.

Admins receive in-app warnings when current usage crosses configured thresholds and a limit notification when the active budget is exhausted. Once a daily or monthly cap is reached, new metered chat and generation calls using that provider are blocked. Workflows that support deferral pause their AI step and are reconsidered after the budget resets or an admin changes the key or cap. These caps do not preflight or block embedding calls, Compensation Research background enrichment, or Hiring Live Review.

Provider Failure

A valid key can still fail because of provider billing, quota, model, or availability problems. Kit shows the affected provider and latest recorded error on the AI settings page. Fix the provider account or replace the key, then retry the affected product action. Do not assume changing the preferred provider replays work that already failed unless the product says it was deferred.

Data Boundary

The preferred provider receives the material needed for chat and generation requests, except for the fixed paths below. Semantic indexing and search always use Google’s gemini-embedding-001 through Kit’s platform Gemini key, even when Anthropic or OpenRouter is preferred and account-keyed. Changing the preference does not reroute embedding data.

Compensation Research Background Enrichment

Compensation Research uses Kit’s fixed Gemini economy model when it cannot finish imported-data cleanup deterministically. These background calls classify listings and extract salary details, match or deduplicate company names, and analyze unmatched job titles to extend role clusters. Depending on the task, Google can receive listing titles and skills, up to 2,000 characters of a listing description, company names, and existing role-cluster names and patterns.

This path uses Kit’s platform Gemini key. It does not use the preferred provider or an account-owned key, preflight AI credits, write to the AI usage ledger, or stop at Kit’s daily and monthly caps. Changing the preferred provider therefore does not reroute this processing, and the AI settings usage table is not a complete cost record for it.

Hiring Live Review

Hiring Live Review always opens Google’s models/gemini-3.1-flash-live-preview with the account’s own Gemini API key. It does not use the preferred provider, refuses to start without that account key, and never falls back to Kit’s platform Gemini key. An Anthropic or OpenRouter preference therefore does not reroute Live Review.

Before the Live connection opens, Kit sends Google a system instruction containing the role title and bounded role description, current stage and screening criteria, submission time, candidate name, application answers, and candidate signals visible to the recruiter. It includes bounded extracted CV text when available; otherwise, for an attached PDF, the browser sends Gemini a two-minute URL for the original file. During the direct browser-to-Google session, the recruiter’s typed messages and any microphone audio they choose to enable go to Gemini. Moving to another candidate can also carry bounded recruiter-authored screening focus from the preceding Live review.

Live Review sits outside Kit’s AI usage ledger and spend caps: Kit does not check AI credits before minting its Live token or record the session’s token and audio usage. Kit’s daily and monthly caps therefore do not stop these sessions; the account’s Google-side billing and quota controls still apply. Kit separately limits Live-token creation to 10 attempts per account member per hour.

Hiring’s AI-authored Slack reminders send the candidate’s name, job title, and waiting duration to the preferred provider. An overdue team-review reminder also sends aggregate yes/no/neutral counts, never scorecard comments. If the configured provider is unavailable or its Kit budget is exhausted, Kit uses deterministic reminder copy instead.

Embedding inputs can include Kit documentation and articles, account knowledge, candidate résumé data and video transcripts, vulnerability report text, Outreach prospect notes, and Compensation Research listings or search text. Review Google as a separate processor for those workflows, alongside the preferred provider. Account-key caps do not apply to these platform-key calls.

Kit records estimated Gemini platform-key usage for some account-scoped embedding workflows, but this is not a complete embedding ledger, and embedding calls are not stopped by the platform-key allowance. Do not rely on AI settings caps to prevent embedding data from being sent.

Review each processor’s terms and your organization’s data rules before enabling AI features. Product features may persist their inputs, sources, drafts, and final records in Kit according to that product’s workflow; the usage table is cost metadata, not a transcript of every request.

Use least privilege around the result as well as the key:

  • Module and record permissions still decide what a member can ask Kit’s AI to use.
  • Generated text remains a draft until the product’s explicit submit, publish, approve, or send action.
  • Check sources and sensitive fields before approving consequential output.
  • Revoke a member’s product access when they no longer need the underlying records; rotating an AI key is not a substitute for access control.

Quick Checklist

  • Confirm which provider may receive chat and generation data
  • Approve Google as the separate processor for semantic indexing, search, and Compensation Research background enrichment
  • Add and validate the account key, or deliberately use the platform key
  • Select the preferred provider
  • For OpenRouter, review all three task-tier models
  • Set Kit caps for chat and generation, provider-side billing alerts, and separate embedding monitoring
  • Run one non-sensitive test and verify the usage attribution
  • Assign an admin to provider failures and key rotation

Type to search...