AI providers

An AI provider is the model service that powers your AI reply suggestions. Every new organization starts with a single provider called Automatic, already set as the default, so AI replies work before you configure anything — it follows whichever model HelpStack currently recommends. You can change or replace it at any time. You configure one or more providers for your organization, choose a model and how it behaves, and pick one as the default. Channels then use either your default provider or one they pick themselves. This guide covers the supported provider types, when to choose each, every configuration option, how the default works, and a checklist for getting your first AI reply working end to end.

You'll find providers at Settings → AI Providers (route /settings/ai-providers). Managing AI providers requires the OWNER or ADMIN role; agents and viewers can generally read them.

The provider types#

When you add a provider, you choose its type:

• OpenAI — OpenAI's hosted models. A good general-purpose default for most teams.
• Anthropic — Anthropic's hosted models (Claude). Another strong general-purpose choice.
• DeepSeek — DeepSeek's hosted models, accessed through an OpenAI-compatible API.
• Google — Google's Gemini models. Fast and inexpensive, a good fit for high-volume chat.
• xAI — xAI's Grok models.
• Qwen — Alibaba's Qwen models. The cheapest option we offer.
• Mistral — Mistral's open-weight models, built in Europe.
• Moonshot — Moonshot's Kimi models, strong on long conversations.
• MiniMax — MiniMax's models.
• Ollama — runs models locally on infrastructure you control. Choose this when you need on-premises inference and no data egress — for example, when customer data must not leave your network. Ollama is local, so it requires a Base URL pointing at your Ollama server, and the model is free-text (you type the model name you've pulled locally).

Which one should you pick?

• Happy with the default? DeepSeek is already set up and working. Nothing to do.
• Want the simplest path with strong quality? Start with OpenAI or Anthropic.
• Want the cheapest running cost? Qwen is the least expensive per reply.
• Must keep data on-premises with no third-party API calls? Use Ollama.

Configuration options#

When adding or editing a provider, you set:

• Type — one of OpenAI, Anthropic, DeepSeek, Google, xAI, Qwen, Mistral, Moonshot, MiniMax, or Ollama (see above).
• Name — a label you choose to identify this provider in the list.
• Model — chosen from a preset list per provider, except Ollama, where the model is free-text (type the name of a model you've pulled locally).
• Temperature — a slider from 0 to 2. Lower values make the AI more focused and consistent; higher values make it more varied/creative. For support replies, lower is usually safer.
• Max tokens — the maximum length of the AI's response.
• Base URL — required for Ollama; optional for the others.

A note on API keys

For the cloud providers, API keys are platform-managed — in the normal setup you do not paste an OpenAI or Anthropic key yourself. Ollama is the exception: because it runs locally, it doesn't use a key but does need its Base URL so HelpStack can reach your server.

The default provider#

Exactly one provider per organization is the default. The default is used by any channel that doesn't pick its own provider.

• Toggle "Set as default" when adding or editing a provider, or use the Set Default action on a provider card.
• Because there's always exactly one default, setting a new default moves the star off the previous one.
• You cannot delete the default provider — set another provider as default first, then delete the old one.

The fallback provider#

If your default provider fails mid-reply, HelpStack answers with a fallback provider instead of leaving the customer without one. Every organization has a fallback, whether or not you set one up. You can see which one you have at the bottom of Settings → AI Providers.

HelpStack picks it in this order:

1. A provider you chose. Turn on "Use as fallback" when editing a provider. Only one provider can be the fallback, the same way only one can be the default.
2. Another provider you already configured, if it's active and on a different vendor.
3. The HelpStack fallback — our own pick, used when you have nothing else suitable. This is what most organizations get, and it needs no setup. We keep more than one, so the pick is always on a different vendor from your default — including when your default happens to be the same vendor we'd otherwise reach for.

Two rules the switch enforces, because a fallback that looks configured but covers nothing is worse than none:

• The default provider can't be its own fallback.
• The fallback must be a different vendor from the default. The failures a fallback exists to survive — an outage, rate limiting, a provider timing out — take all of one vendor's models at once, so falling back from one OpenAI model to another is the same outage twice. This is why the "Use as fallback" switch doesn't appear when you edit your default provider, or one that shares its vendor.

Your system prompt, temperature and max tokens carry over to the fallback. Only the vendor and model change, so a fallback reply doesn't suddenly answer in a different voice.

A fallback is only used for failures where a different vendor could plausibly succeed. Being out of credits, a rejected API key, or a malformed request are not retried elsewhere — retrying those would spend money reproducing the same failure and hide its cause.

Reading the provider cards#

Each provider in the list shows:

• Name
• Type badge (OpenAI / Anthropic / DeepSeek / Google / xAI / Qwen / Mistral / Moonshot / MiniMax / Ollama)
• Default star (if it's the default), or a Fallback badge (if it's the fallback)
• Model
• Temperature
• Max tokens
• Active / Inactive status
• Actions: Set Default, Edit, Delete (you can't delete the default)

Where the system prompt lives#

The system prompt — the AI's persona and instructions (its tone, what it should and shouldn't do) — is not set on the provider. It's configured per-channel, in the channel's AI & Knowledge tab. This lets you give different channels different personalities while sharing the same provider. See Channels.

Also note: the customer's email/identifier is automatically made available to the AI in its prompt, so it can help with account-specific questions without you wiring that in manually.

Getting your first AI reply working#

A working AI reply needs three things in place. Use this checklist:

1. Configure a provider. Your account already has a DeepSeek provider set as the default, so you can skip this step. To use a different one, add any of the provider types above at Settings → AI Providers, pick a model and temperature, and set it as the default (or plan to assign it to a specific channel). For Ollama, set the Base URL.
2. Attach a knowledge base to the channel. So replies are grounded in your real content, attach knowledge-base groups to the channel under Settings → Channels → channel → AI & Knowledge. See Knowledge base and Channels.
3. Enable AI replies on the channel. Turn on AI replies for the channel (and set its system prompt in the same AI & Knowledge tab). See AI replies.

With those three done, send a test message on the channel and confirm the AI drafts a reply grounded in your knowledge base.

Related#

• Channels — assign a provider per channel and set the per-channel system prompt.
• AI replies — how AI reply suggestions work and how to enable them.
• Knowledge base — ground replies in your own content.
• Agent tools (overview) — give the AI custom abilities while it drafts.
• Getting started — initial organization and channel setup.
• Glossary — definitions of temperature, tokens, system prompt, and more.