The “Models” screen decides which language models your agents use. The core is
a priority-ordered provider chain: AiHummer walks the list top to bottom until
it gets a working answer, so you can put your primary provider first and fallbacks
below.
[!NOTE]
A real provider is not required: until you connect one, replies come from a
deterministic mock — enough to test channels, routing and agents at no cost.
Connect a model when you want “real” answers.
Connect your first provider in 4 steps
Open the “Models” screen and click the add-provider button — a modal opens
(the same one used in the setup wizard).
Pick a provider from the branded list (it cannot be changed after
creation). Russian providers come first, then Codex, then international ones:
Russian:YandexGPT (Yandex Cloud), GigaChat (Sber),
Cloud.ru (Evolution), MTS AI (Cotype). The preset fills the correct
endpoint for you — just add your key and pick a model.
codex — models via a ChatGPT subscription with no separate API key
(see below).
International: OpenAI, Anthropic (Claude), Google Gemini, DeepSeek,
xAI (Grok), Mistral, Groq and more — plus openai-compatible for any
service with an OpenAI-compatible API and local servers (Ollama, LM Studio,
vLLM).
Most presets use an OpenAI-compatible endpoint (the address is filled in for
you); YandexGPT and GigaChat have their own built-in authentication, so
no separate base_url is needed.
Fill in the fields: a Name (any label, for yourself), the Base URL
(auto-filled by the preset) and the API key. Click
Test — AiHummer checks connectivity.
Pick a model: you can type the name by hand first (with hints), and after a
successful Test the model list is fetched from the provider’s /models.
Tick “enabled” and save.
What you’ll see: the new provider appears as a card with a priority number, type,
model, a “key set” badge and an “enabled” status. Agents now answer through it
instead of the mock.
What goes where: provider examples
Pick a provider from the list — the preset fills the address; you just add your
key (find it in the provider’s own docs) and pick a model. The table below
is a guide (Russian providers at the top).
Provider
What to set
YandexGPT (Yandex Cloud)
Yandex Cloud key + a YandexGPT model (built-in auth)
GigaChat (Sber)
GigaChat key + a GigaChat model (built-in auth)
Cloud.ru (Evolution)
Cloud.ru key + a model
MTS AI (Cotype)
MTS AI key + a Cotype model
Codex (ChatGPT subscription)
no key — device authorization (see below)
OpenAI
OpenAI key + a model (e.g. a GPT-family model)
Anthropic
Anthropic key + a Claude model
Google Gemini
Google key + a Gemini model
DeepSeek
DeepSeek key + a DeepSeek model
xAI (Grok), Mistral, Groq, etc.
the provider’s key + its model
OpenAI-compatible / local server
your service’s base_url (Ollama/LM Studio/vLLM), key if needed
How many providers, and how fallback works
The provider chain is tried in order: if the top provider is unavailable or
returns an error, AiHummer moves to the next. So:
put your primary provider first (the ↑/↓ arrows reorder);
add a fallback below — for failures or rate limits on the primary;
you can combine a cloud provider with a local one as a backup.
Codex (ChatGPT subscription)
The codex type hides base_url and key and shows a device-auth connection
panel: Connect / Disconnect / Refresh. It lets you use models via a ChatGPT
subscription with no separate API key:
Add a provider with the codex type.
Click Connect and complete device authorization per the on-screen prompt.
Done — the provider runs through your subscription; Refresh renews access,
Disconnect turns it off.
BYOK and LLM settings
BYOK (bring your own key) — a per-workspace override: provider, model,
api_key, base_url; “Save” and “Clear” buttons. Details:
BYOK providers.
LLM settings — part of the settings catalog pinned to this screen
(temperature, timeouts, etc.).
Limiting what agents can use
By default an agent may use the chain’s default model. To limit the models
available to agents, set the list on
Instructions → allowed models.
If it didn’t work
Test fails — check that the Base URL points at the provider’s
OpenAI-compatible endpoint (not a sign-in page) and that the API key is
copied in full and active. Different providers use different addresses — check
their docs.
The model list is empty — models are fetched from /models only after a
successful Test; before that, type the model name by hand.
The agent still answers like the mock — make sure the provider is enabled
(ticked) and sits above disabled ones in the chain.
Rate-limit/quota errors — add a fallback provider lower in priority so
requests switch to it when the primary is exhausted.
Wrong type picked — the type can’t be changed after creation; delete the
provider and recreate it with the correct type.
The **"Models"** screen decides which language models your agents use. The core is
a **priority-ordered provider chain**: AiHummer walks the list top to bottom until
it gets a working answer, so you can put your primary provider first and fallbacks
below.
> [!NOTE]
> A real provider is **not required**: until you connect one, replies come from a
> deterministic mock — enough to test channels, routing and agents at no cost.
> Connect a model when you want "real" answers.
## Connect your first provider in 4 steps
1. Open the **"Models"** screen and click the add-provider button — a modal opens
(the same one used in the setup wizard).
2. Pick a **provider** from the branded list (it cannot be changed after
creation). Russian providers come **first**, then Codex, then international ones:
- **Russian:** **YandexGPT** (Yandex Cloud), **GigaChat** (Sber),
**Cloud.ru** (Evolution), **MTS AI** (Cotype). The preset fills the correct
endpoint for you — just add your key and pick a model.
- **`codex`** — models via a **ChatGPT subscription** with no separate API key
(see below).
- **International:** OpenAI, Anthropic (Claude), Google Gemini, DeepSeek,
xAI (Grok), Mistral, Groq and more — plus **`openai-compatible`** for any
service with an OpenAI-compatible API and local servers (Ollama, LM Studio,
vLLM).
Most presets use an OpenAI-compatible endpoint (the address is filled in for
you); **YandexGPT** and **GigaChat** have their own built-in authentication, so
no separate base_url is needed.
3. Fill in the fields: a **Name** (any label, for yourself), the **Base URL**
(auto-filled by the preset) and the **API key**. Click
**Test** — AiHummer checks connectivity.
4. Pick a **model**: you can type the name by hand first (with hints), and after a
successful Test the model list is **fetched** from the provider's `/models`.
Tick "enabled" and save.
*What you'll see:* the new provider appears as a card with a priority number, type,
model, a "key set" badge and an "enabled" status. Agents now answer through it
instead of the mock.
### What goes where: provider examples
Pick a provider from the list — the preset fills the address; you just add your
**key** (find it in the provider's own docs) and pick a **model**. The table below
is a guide (Russian providers at the top).
| Provider | What to set |
|---|---|
| YandexGPT (Yandex Cloud) | Yandex Cloud key + a YandexGPT model (built-in auth) |
| GigaChat (Sber) | GigaChat key + a GigaChat model (built-in auth) |
| Cloud.ru (Evolution) | Cloud.ru key + a model |
| MTS AI (Cotype) | MTS AI key + a Cotype model |
| Codex (ChatGPT subscription) | no key — device authorization (see below) |
| OpenAI | OpenAI key + a model (e.g. a GPT-family model) |
| Anthropic | Anthropic key + a Claude model |
| Google Gemini | Google key + a Gemini model |
| DeepSeek | DeepSeek key + a DeepSeek model |
| xAI (Grok), Mistral, Groq, etc. | the provider's key + its model |
| OpenAI-compatible / local server | your service's base_url (Ollama/LM Studio/vLLM), key if needed |
## How many providers, and how fallback works
The provider chain is **tried in order**: if the top provider is unavailable or
returns an error, AiHummer moves to the next. So:
- put your primary provider **first** (the ↑/↓ arrows reorder);
- add a **fallback** below — for failures or rate limits on the primary;
- you can combine a cloud provider with a local one as a backup.
## Codex (ChatGPT subscription)
The **`codex`** type hides base_url and key and shows a device-auth connection
panel: **Connect / Disconnect / Refresh**. It lets you use models via a ChatGPT
subscription with no separate API key:
1. Add a provider with the `codex` type.
2. Click **Connect** and complete device authorization per the on-screen prompt.
3. Done — the provider runs through your subscription; **Refresh** renews access,
**Disconnect** turns it off.
## BYOK and LLM settings
- **BYOK** (bring your own key) — a per-workspace override: provider, model,
api_key, base_url; "Save" and "Clear" buttons. Details:
[BYOK providers](/en/v1.0/access/byok-providers).
- **LLM settings** — part of the settings catalog pinned to this screen
(temperature, timeouts, etc.).
## Limiting what agents can use
By default an agent may use the chain's default model. To limit the models
available to agents, set the list on
[Instructions → allowed models](/en/v1.0/webui/skills-prompts#instructions).
## If it didn't work
- **Test fails** — check that the **Base URL** points at the provider's
OpenAI-compatible endpoint (not a sign-in page) and that the **API key** is
copied in full and active. Different providers use different addresses — check
their docs.
- **The model list is empty** — models are fetched from `/models` only after a
successful Test; before that, type the model name by hand.
- **The agent still answers like the mock** — make sure the provider is **enabled**
(ticked) and sits **above** disabled ones in the chain.
- **Rate-limit/quota errors** — add a fallback provider lower in priority so
requests switch to it when the primary is exhausted.
- **Wrong type picked** — the type can't be changed after creation; delete the
provider and recreate it with the correct type.
## Next
- [Routing](/en/v1.0/concepts/routing) — how the model and agent are chosen.
- [BYOK providers](/en/v1.0/access/byok-providers) — per-workspace keys.
- [Agents](/en/v1.0/webui/agents) — give an agent a dedicated model.
