Skip to content

LLM Model Configs

Use the LLMs tab to create chat model configs. Each config points to a provider-specific chat connection and stores the model identifier plus generation parameters.

Supported chat providers are Ollama, OpenAI, Anthropic, OpenRouter, Google GenAI, Mistral AI, Cohere, Perplexity, Groq, DeepSeek, xAI, and Cloudflare. The UI shows provider-specific fields after you choose the provider, and only connections with the matching provider type are available for that config.

Common fields

Field Description
name Display name for this LLM profile.
model Provider model identifier, for example an Ollama model name or OpenAI-compatible model id.
temperature Sampling temperature. Higher values are more varied; lower values are more deterministic.
context_window Context size used when composing prompts.
seed Optional seed for providers that support deterministic sampling.
reasoning Optional reasoning control. It may be true, false, or a level such as low, medium, or high, depending on provider support.
stop_tokens Optional comma-separated stop tokens.

Complex provider parameters, such as model_kwargs, custom headers, response schemas, penalty configs, and safety settings, are entered as JSON in the frontend. Leave optional fields empty to use the provider default.

Provider-specific fields

Provider Additional configuration
Ollama Mirostat controls, prediction length, repeat penalty controls, model validation on init, GPU/thread options, logprobs, top_k, top_p, output format, keep_alive, and sync/async client kwargs.
OpenAI model_kwargs, organization, proxy, request timeout, retries, penalties, logprobs, streaming, top_p, completion token limit, reasoning effort, verbosity, tiktoken model name, default headers/query, socket options, stream timeout, extra body, response headers, disabled params, Responses API context/include/service-tier/store/truncation options, and Responses API conversation flags.
Anthropic model_kwargs, max tokens, timeout, retries, top_p, top_k, thinking config, output config, usage streaming, streaming, default headers, beta flags, service tier, MCP servers, container, and inference geography.
OpenRouter OpenAI-compatible controls, plus any OpenRouter-specific request body values through model_kwargs or extra_body.
Google GenAI model_kwargs, max output tokens, top_p, top_k, candidate count, retries, timeout, safety settings, response MIME type, response schema, cached content, thinking budget, thought inclusion, transport, and client options.
Mistral AI model_kwargs, max tokens, top_p, random seed, safe mode, streaming, endpoint override, timeout, retries, and concurrent request limit.
Cohere model_kwargs, preamble, streaming, user agent, and request timeout.
Perplexity OpenAI-compatible controls, with unsupported parameters ignored or rejected by the provider depending on the model/API.
Groq OpenAI-compatible controls plus max tokens, reasoning format, response format, and parallel tool-call control.
DeepSeek OpenAI-compatible controls, with provider-specific values available through model_kwargs or extra_body.
xAI OpenAI-compatible controls plus live search parameters.
Cloudflare model_kwargs, account ID, endpoint format, AI Gateway slug, max tokens, top_p, top_k, and streaming.

Assignments

After creating LLM configs, assign them to a world or simulation from the world editor or simulation detail page.

LLM-capable simulation components include input_interpreter, ooc_handler, action_validator, character_simulator, scene_coordinator, state_committer, memory_summarizer, perspective_resolver, trigger_evaluator, action_suggester, and narrator.

Image-related chat components include character_image_generator, character_portrait_image_generator, location_image_generator, item_image_generator, scene_image_generator, and turn_image_trigger.