LLM Providers

Molf Assistant uses the Vercel AI SDK (ai package) for all LLM interactions. It supports 16 bundled providers and can be extended with custom provider definitions.

Bundled Providers

Provider	API Key Variable
google	GEMINI_API_KEY
anthropic	ANTHROPIC_API_KEY
openai	OPENAI_API_KEY
openai-compatible	OPENAI_API_KEY
xai	XAI_API_KEY
mistral	MISTRAL_API_KEY
groq	GROQ_API_KEY
deepinfra	DEEPINFRA_API_KEY
cerebras	CEREBRAS_API_KEY
cohere	COHERE_API_KEY
togetherai	TOGETHER_AI_API_KEY
perplexity	PERPLEXITY_API_KEY
amazon-bedrock	AWS_ACCESS_KEY_ID
google-vertex	GOOGLE_APPLICATION_CREDENTIALS
azure	AZURE_OPENAI_API_KEY
openrouter	OPENROUTER_API_KEY

Providers are auto-detected based on the presence of their API key environment variable. Only providers with a detected key are available.

To enable all providers with detected keys regardless of the config file:

MOLF_ENABLE_ALL_PROVIDERS=1 pnpm dev:server

Or in molf.yaml:

enable_all_providers: true

To enable only specific providers:

enabled_providers:
  - google
  - anthropic

Model Resolution

Models are specified in provider/model-name format (e.g., google/gemini-2.5-flash or anthropic/claude-sonnet-4-20250514).

The model for a given prompt is resolved in this priority order:

1. Prompt-level -- model specified in the agent.prompt call
2. Workspace config -- model set via workspace.setConfig
3. Server default -- set via MOLF_DEFAULT_MODEL env var or model in molf.yaml

Provider Initialization

The provider registry follows a 7-step pipeline on startup:

1. Parse the model ID into provider and model name
2. Load the model catalog (from models.dev or bundled snapshot)
3. Build the list of allowed providers from config
4. Detect API keys in environment variables
5. Create SDK instances for detected providers
6. Resolve the requested language model
7. Return the provider state and resolved model

models.dev Catalog

The server fetches an up-to-date model catalog from models.dev to provide model metadata and validation.

Setting	Value
Catalog URL	https://models.dev/api/models
Fetch timeout	5s
Refresh interval	60 min

The fallback chain when fetching models:

1. In-memory cache
2. Disk cache (persisted between restarts)
3. Bundled snapshot (shipped with the package)
4. Live fetch from models.dev API

To disable the models.dev catalog entirely:

MODELS_DEV_DISABLE=1 pnpm dev:server

Custom Providers

Define custom providers in molf.yaml under the providers key:

providers:
  my-provider:
    # Custom provider configuration

Custom providers are registered alongside the bundled ones and can be referenced using the same provider/model format.

Provider-Specific Behavior

Some providers have special handling:

• Anthropic -- enables the interleaved-thinking beta header automatically
• OpenAI -- GPT-5 and later models use the responses API instead of the chat completions API

Behavior Configuration

The behavior section in molf.yaml controls agent behavior:

behavior:
  temperature: 0.7
  contextPruning: true

Field	Description	Default
temperature	LLM sampling temperature	Provider default
contextPruning	Enable automatic context pruning	true

The BehaviorConfig also supports systemPrompt and maxSteps (default: 10) when set programmatically.

See Also

• Configuration -- full configuration reference including provider API keys
• Sessions -- how model resolution works per-prompt
• Architecture -- provider registry in the package graph