LLM Providers
Choose and configure the AI backend that powers your agents
What It Does
TorlyAI supports multiple LLM providers, so you can choose the AI backend that best fits your needs. Each workspace can use a different provider, and sessions lock to their provider after the first message for consistency.
Supported Providers
Anthropic Claude
RecommendedThe recommended provider for TorlyAI. Claude excels at structured business writing, nuanced analysis, and following complex instructions. All TorlyAI skills are optimised for Claude.
OpenAI
GPT-4 and later models. Compatible with TorlyAI skills, though some specialised outputs may vary from Claude-optimised examples.
GitHub Copilot
Uses your existing GitHub Copilot subscription. Good for teams already using GitHub Copilot who want a unified billing model.
OpenRouter
A gateway to multiple model providers through a single API key. Supports dozens of models including Claude, GPT-4, Llama, and Mixtral.
Ollama
LocalRun models locally on your own hardware. Fully offline with no data leaving your machine. Performance depends on your hardware and model choice.
Per-Workspace Selection
Each workspace can use a different provider. This is configured in the workspace settings and lets you:
- Use Claude for your primary business plan workspace
- Use Ollama for a local-only workspace with sensitive data
- Experiment with different providers in separate workspaces
Session Provider Lock
Once a session sends its first message, it locks to the provider that was active at that moment. This ensures the entire conversation uses the same model, preventing inconsistencies from provider switching mid-conversation.
Changing your workspace provider only affects new sessions. Existing sessions continue using their locked provider.
Secondary LLM Calls
Agents can make secondary LLM calls for subtasks using the call_llm tool. This enables an agent to delegate smaller tasks (like parsing a document or generating structured JSON) without interrupting the main conversation flow.
File Attachments
Attach files to secondary calls for analysis or summarisation.
Structured Output
Request JSON output for structured data extraction and parsing.