Providers & Models

TomoriBot doesn’t have a built-in AI model, you connect one from a provider. A provider is an AI service (Google Gemini, OpenRouter, NovelAI, a local endpoint, …), and a model is a specific model on that provider. You need at least one provider to use her at all.

API Keys

Add a provider key during first-time setup with /setup, or later from /providers by choosing Add New Provider. Keys are encrypted at rest, so no one, including server admins, can read them back.

/setup asks how replies should reach a model before anything else, and the answer decides what it collects:

Mode	What it collects
AI Provider (Recommended)	A provider from the catalog plus its API key, validated and encrypted as a draft.
Custom Endpoint (Advanced)	The endpoint’s connection and one text model, registered inside the wizard. See Custom Endpoints.
User BYOK (guilds only)	Nothing: the workspace keeps no provider of its own, so members must supply personal ones.

Nothing is written until Finish Setup, so an abandoned or expired wizard leaves the workspace’s existing provider rows alone. To replace a key that is already stored, use /providers, because /setup refuses to run on a workspace that is already configured.

Each provider has its own key-generation steps. Run /help, choose Setup, then Step 1: Get an API Key, and pick your provider for the exact walkthrough, or use these starting points:

Provider	Notes	Get a key
Google Gemini	Free tier, runs every feature. Recommended first setup.	AI Studio
OpenRouter	One key, many models (some free).	OpenRouter keys
NovelAI	Subscription; uncensored storytelling/roleplay (text only).	NovelAI
DeepSeek	Pay-as-you-go reasoning models.	DeepSeek
NVIDIA NIM	Hosted text, embeddings, and image.	NVIDIA Build
Anthropic	Claude models via the API (not Claude Code).	—
Z.ai	GLM family. ⚠️ ToS restricts usage to coding/agent scenarios.	Z.ai
Vertex AI	Google Cloud via `gcloud` ADC (best for locally-run/dev setups).	see below
Vertex AI Express	Google Cloud API-key BYOK (Preview, Gemini subset).	Express Mode
Custom	Any OpenAI-compatible endpoint (Ollama, vLLM, LiteLLM, …).	see Custom Endpoints

Vertex AI authenticates with Application Default Credentials rather than a stored secret. For local hosting, ADC can come from gcloud; hosted deployments should use a workload identity or service account. An AI Studio API key alone does not authenticate full Vertex AI. The selected project must have billing and the Vertex AI API enabled, and the host identity needs Vertex access. The setup guide is available from Google Vertex AI on the API Keys page in /help.

Google-backed provider setup validates credentials through the authenticated model-listing endpoint. It does not generate text or depend on whichever chat model is currently marked as the catalog default, so a retired default cannot prevent a valid credential from being saved.

Optional: Brave Search key

Brave Search is separate from your AI provider and only enhances web search (adds image, video, and news search). Set it with /providers. ⚠️ Brave includes $5/month free credit, so set a $5 usage limit in the Brave dashboard to avoid charges.

Choosing Models

/providers manages server credentials, model catalogs, and endpoint registrations, while /config > Models > Switch Models selects the shared capability assignments every member of this server uses. Both need the required server permission. Individual members manage their own credentials and model catalogs with /personal providers, then select personal models in /personal config. Personal settings follow them across every server where they use TomoriBot. See Personalization for that side.

The panels are titled Server Providers and Personal Providers so their ownership remains visible after the command interaction opens.

After a provider is set, use /config > Models > Switch Models to choose the shared capability assignments. The six ordinary slots select model entries from provider catalogs:

/config > Models > Switch Models: the main chat model
/config > Models > Switch Models: a vision model (for reading images when the chat model can’t)
/config > Models > Switch Models: embeddings for the document knowledge base
/config > Models > Switch Models: standard image generation (see Image Generation)
/config > Models > Switch Models: NovelAI image generation
/config > Models > Switch Models: video generation
/config > Models > Switch Models: text-to-speech (TTS) endpoint
/config > Models > Switch Models: speech-to-text (STT) endpoint

The first six entries choose model catalog records. The TTS and STT slots choose workspace-scoped endpoints instead, so they activate the selected endpoint rather than writing a model column. Register and edit those endpoints in /providers; its endpoint activation control still works. /personal config retains six personal model-routing slots and does not add personal TTS/STT endpoint selectors.

You can also manage this server’s backup keys for automatic failover and load balancing with /providers.

Custom Endpoints

Custom endpoints let you register self-hosted or proxy-backed services (Ollama, LM Studio, LiteLLM, vLLM, ComfyUI, local TTS/STT) as labeled provider bundles.

Server scope: open /providers for workspace endpoint registration and editing.
Personal scope: open /personal providers for personal model catalogs (just you: see Personalization). Personal speech endpoints are not selected from /personal config.

A label is the user-facing menu name and groups capabilities under one bundle when they share one endpoint URL. It is never sent to the remote endpoint. Capabilities served from different URLs need distinct labels. Choose Add New Custom Endpoint, select the API compatibility, and save the connection. Saving prepares the capabilities supported by that protocol without registering any models. Then select the new endpoint and use its model dropdown to register an exact model code and capability. Adding a model activates it for that capability. Use the same dropdown to attach more models or edit a workspace-added registration. Text models declare their own capabilities in that form, and image models declare which request modes they support.

For TTS and STT, register the endpoint and its models in /providers, then choose and activate the endpoint in /config > Models > Switch Models. Those speech slots select an endpoint rather than a model catalog entry. /providers remains the endpoint registration, model setup, and editing surface.

API compatibility determines the request paths and payloads the service implements, so it also determines which capability slots the connection prepares. Registering exact models for those slots is a separate step, and the protocol cannot be inferred reliably from the endpoint URL.

/setup’s Custom Endpoint (Advanced) mode performs the same two steps inside the wizard: Configure Connection saves the API compatibility, label, URL, and optional auth token behind a reachability check, and Configure Text Model registers the exact text model and its capability declarations. The model button stays disabled until a connection validates, and re-saving the connection clears the model declaration because the declarations depend on the API compatibility. The wizard creates the connection, saved provider, model, and active-model rows together when you press Finish Setup, so it never leaves a connection that has no usable text model. It registers text models only; image, video, TTS, and STT capabilities are still registered in /providers.

For full walkthroughs of running the servers, see:

Setup: Local LLM: Ollama, KoboldCPP, LM Studio, vLLM, LiteLLM.
Setup: ComfyUI: local image/video generation.
Setup: ChatMock: ChatGPT account / Codex CLI.

Supported Providers

If you don’t have the hardware to host your own models, TomoriBot supports a wide range of services. Not every feature is available on every provider.

LLM Providers

Provider	Streaming	Tool Calling	Image Input	Embeddings	Notes
Google Gemini	✅	✅	✅	✅	Free models available
OpenRouter	✅	✅	✅	✅	Free models available
Anthropic (API)	✅	✅	✅	–	Not Claude Code
NovelAI	✅	✅	–	–	Only GLM 4.6 can use tools
NVIDIA NIM	✅	✅	✅	✅	Free models available
DeepSeek	✅	✅	–	–	–
Z.ai	✅	✅	✅	–	Free models; ⚠️ ToS = coding/agent use only
Z.ai Coding	✅	✅	–	–	Subscription plan
Google Vertex AI	✅	✅	✅	✅	Includes ‘free’ Express version
Codex CLI (via ChatMock)	✅	✅	✅	–	Setup

Image Generation

Provider	Text-to-Image	Image-to-Image	Inpainting	Notes
Google	✅	✅	–	–
OpenRouter	✅	✅	–	–
NovelAI	✅	✅	✅	Can combine with other providers
NVIDIA	✅	–	–	Text-to-image only; reference images are ignored
Z.ai	✅	–	–	–

These are the defaults a provider’s image models start from, and NovelAI runs through its own pipeline rather than this table. Registering an image model through /providers lets you declare that model’s own modes, which is how you enable inpainting on a ComfyUI workflow or on a provider model whose API supports masked editing. A model you never declare keeps following the defaults above, so a later correction to them reaches it automatically. Declare only what the model really does: Tomori offers the tool exactly the modes you tick, and a mode the API rejects becomes a failed generation.

Video Generation

Provider	Text-to-Video	Image-to-Video	Notes
Google	✅	✅	Async polling workflow
OpenRouter	✅	✅	Async polling workflow
Z.ai	✅	✅	Async polling workflow

Voice & Audio

Provider	Text-to-Speech	Speech-to-Text
ElevenLabs	✅	✅

Local voice engines are covered under Self-Hosting. For the built-in web search and URL-fetch engines, see Tools & Extensions.