Skip to main content

Models and Providers

TARX uses LiteLLM as its LLM routing layer, giving you a single, unified way to connect every major provider. You configure a key, pick a model from the live list, and TARX handles the protocol differences behind the scenes.

This page is about how to integrate and use models in TARX — not a catalog of third-party models or pricing. Model availability, capabilities, and cost are set by each provider and change constantly; TARX always reflects what your key can actually do by listing models live.

A single backend provider registry is the source of truth for every supported provider and its routing rules. The Agent Editor and the key form both read it from GET /api/v1/llm-providers, so what the UI offers always matches what the platform supports.


Supported Providers

ProviderKey prefix / sourceNeeds base URL?Tool callingOfficial model reference
Anthropicsk-ant-…Nodocs.anthropic.com/models
OpenAIsk-…Noplatform.openai.com/docs/models
Google AI (Gemini)AIza…Noai.google.dev/gemini-api/docs/models
MistralMistral API keyNodocs.mistral.ai/getting-started/models
HuggingFacehf_…Nohuggingface.co/models
Azure AI FoundryAzure key✅ (resource endpoint)your deployments
OpenAI-compatibleprovider key✅ (base URL)provider's /v1/models

The OpenAI-compatible provider is the generic option for any service implementing the OpenAI API spec — Groq, Together, OpenRouter, Fireworks, Perplexity, Ollama, vLLM, and more. See the full list and connection steps in BYOK: API Keys → OpenAI-Compatible Endpoints.

Tool calling matters for capabilities

Agent capabilities (web_search, web_scraper) and MCP tools rely on the model's tool/function-calling support. Most modern models support it; some smaller open-source models (and HuggingFace inference) may not. If an agent never calls its tools, check that the model supports tool calling.


Where the Model List Comes From (Live per Key)

You don't type a model string from memory. When you select a project LLM key in the Agent Editor, TARX lists the models that key can actually use. The editor always tells you the source of the list ("N models from this key" vs. "common models shown") and always lets you type a model id by hand, so the experience is the same for every provider: live list → curated fallback → free-text by documented id.

ProviderListing sourceReliability
OpenAI · OpenAI-compatible · Mistralthe provider's GET /v1/models with your keyHigh — live & self-updating
Google AI (Gemini)the Gemini models APIHigh — live & self-updating
Anthropicthe Anthropic models APIHigh — live & self-updating
HuggingFacepopular warm text-generation models from the HF Hub (not key-scoped)Medium — dynamic; curated fallback
Azure AI Foundryyour resource's deploymentsBest-effort — see the Azure rule below

If the live call fails (offline, quota, network), TARX falls back to a small curated list and lets you type a model id by hand — each provider's field links to its official model-id reference.

Azure: use the resource endpoint, not a deployment URL

Azure can only enumerate deployments from the resource-level endpoint + key shown on the Foundry Overview page (e.g. https://<resource>.cognitiveservices.azure.com/). A single deployment's URL (…/openai/deployments/<name>/chat/completions?api-version=…) cannot list other deployments — the key form rejects it. You can only use models you have actually deployed in Azure; if listing isn't available, type the deployment name (the azure/ prefix is added for you).


How Routing Works

Whatever model id you pick or type, TARX routes it correctly for the provider — you don't add prefixes yourself:

ProviderRouting appliedExample id you'd see/enter
OpenAIas-isgpt-4o, gpt-4o-mini
Anthropicas-isclaude-sonnet-4-6
Google AIgemini/…gemini-2.5-pro
Mistralmistral/…mistral-large-latest
HuggingFacehuggingface/…meta-llama/Llama-3.3-70B-Instruct
Azure AI Foundryazure/<deployment> + endpoint + api-versionmy-gpt4o-deployment
OpenAI-compatibleopenai/<model> + base URLllama-3.3-70b-versatile

Model / Key Pairing

Your API key determines which provider's models you can use — you can't call a GPT model with an Anthropic key. Because the model list is sourced from the key itself, the pairing is enforced for you. When you save an agent, TARX runs a quick live validation call, so a broken model/key pairing surfaces at save time instead of mid-workflow.

If your key is for…You can use…
OpenAIgpt-*, o1-*, o3-*, o4-* models
Anthropicclaude-* models
Google AIGemini models
MistralMistral models
HuggingFaceHuggingFace-hosted models
Azure AI Foundryyour deployments (listed live)
OpenAI-compatiblewhatever the base URL's /v1/models returns

Choosing a Model

TARX is model-agnostic — the "best" model depends on your task, your provider account, and your budget. A few practical guidelines:

  • Simple tasks (classify, extract, short summaries): a small/fast model keeps cost and latency low.
  • Complex reasoning, planning, long-form writing: a larger flagship model.
  • Very long inputs: pick a model with a large context window.
  • High volume / speed-critical: fast inference providers (e.g. an OpenAI-compatible service like Groq) shine.

You can always create two agents with the same prompt and different models to compare — duplicate an agent and swap the model.


Keeping Up with New Models

LLM providers release new models frequently. TARX doesn't hardcode model lists — it lists them live from your key, so new models appear automatically the moment the provider exposes them. When live listing isn't available you can type a documented model id by hand; each field links to the provider's official reference: