LLM configuration reference

An LLM config tells an agent which model to call and how to authenticate. Each agent references one by name (llmConfigName). The same configs ship with both the Admin UI and the CLI, under src/main/resources/initial-data/llm-configs/.

You select a model per agent by setting its llmConfigName to one of these.

tip

For managing configs in the Admin UI (create / edit / test, and how API keys resolve from environment variables), see Admin UI → LLM Configs.

Supported providers

Config	Provider	Default model	Auth / endpoint env vars
`claude-default`	Anthropic	`claude-sonnet-4-6`	`ANTHROPIC_API_KEY`
`claude-haiku-default`	Anthropic	`claude-haiku-4-5`	`ANTHROPIC_API_KEY`
`openai-default`	OpenAI	`gpt-5.4-mini`	`OPENAI_API_KEY`
`azure-openai-default`	Azure OpenAI	deployment `gpt-4o`	`AZURE_OPENAI_API_KEY`, `AZURE_OPENAI_ENDPOINT`, `AZURE_OPENAI_DEPLOYMENT`
`gemini-default`	Google Gemini	`gemini-2.0-flash`	`GEMINI_API_KEY`
`lm-studio-default`	LM Studio (local)	`openai/gpt-oss-120b`	none — local server at `http://localhost:1234`
`gemma-reader`	LM Studio (local)	`google/gemma-4-e4b`	none — a small local model for page-reading sub-agents, `GEMMA_MODEL` to override
`agent-default`	alias → `openai-default`	—	whatever the target needs; the default every bundled agent uses
`embeddings`	LM Studio (local), `EMBEDDING`	`text-embedding-nomic-embed-text-v1.5`	none — `EMBEDDING_BASE_URL`, `EMBEDDING_MODEL` to override
`openai-embeddings`	OpenAI, `EMBEDDING`	`text-embedding-3-small`	`OPENAI_API_KEY`, `OPENAI_EMBEDDING_MODEL`

agent-default is the default

Every bundled agent references agent-default. It is an alias, not a provider config of its own: out of the box it delegates to openai-default, so the agents work as soon as OPENAI_API_KEY is set. To move all of them to another provider at once, repoint the alias — edit delegatesTo in agent-default.json or in the Admin UI (for example to claude-default, or to lm-studio-default for a fully local setup). To move a single agent, change its llmConfigName instead.

Anatomy of a config

llm-configs/claude-default.json
{
  "name": "claude-default",
  "provider": "ANTHROPIC",
  "model": "${CLAUDE_MODEL:claude-sonnet-4-6}",
  "baseUrl": "https://api.anthropic.com",
  "apiKey": "${ANTHROPIC_API_KEY}",
  "contextWindowTokens": 200000,
  "capabilities": ["TOOL_CALLING", "VISION", "DOCUMENTS"],
  "additionalParams": {
    "thinking": "adaptive",
    "effort": "high"
  },
  "rateLimit": { "maxConcurrentRequests": 3 },
  "retry": {
    "enabled": true,
    "maxAttempts": 5,
    "baseBackoffMillis": 2000,
    "maxBackoffMillis": 30000
  },
  "fallbackModels": ["openai-default", "gemini-default"]
}

Field	Meaning
`provider`	`ANTHROPIC`, `OPENAI`, `AZURE_OPENAI`, `GOOGLE_GEMINI`, `LM_STUDIO`, `GROQ`, `OLLAMA`, `MISTRAL`, `DEEPSEEK`, `XAI` (Grok), `MOONSHOT` (Kimi), `TOGETHER`, `OPENROUTER`, `PERPLEXITY` or `FIREWORKS`
`model`	Model id. `${VAR:default}` reads an env var with a fallback
`baseUrl`	API endpoint. Optional: each provider knows its own (`LlmProvider.defaultBaseUrl()`), so set it only for a proxy, a local server, or an Azure resource
`apiKey`	API key, usually injected from an env var
`contextWindowTokens`	Token budget used to size the working-memory window
`capabilities`	What the model reads and does — `TOOL_CALLING`, `VISION`, `DOCUMENTS`, `AUDIO_INPUT`. Vision and documents decide whether an image or PDF reaches the model as content or as a placeholder. Omitted: the provider's default applies
`additionalParams`	Provider-specific options — Anthropic `thinking` / `effort`, `reasoning_effort` on every OpenAI-compatible provider
`rateLimit`	Optional — cap concurrent requests
`retry`	Optional — automatic retry with backoff
`fallbackModels`	Optional — other configs, by name, to switch to when this one is rate-limited (see rate limits and fallbacks)

What a config costs is not part of it: prices are their own entity, one per config, model and validity period — a price applies only to calls the config served with that model, so a changed model (or a ${VAR} that now resolves to another one) never inherits the old model's rates — maintained in the admin UI's Pricing section and stored beside the configs (system/llm-prices/, or mc_llm_price on Postgres).

Setting API keys

Export the relevant variable before starting the Admin UI or CLI. For example:

export ANTHROPIC_API_KEY=sk-ant-...
# or
export OPENAI_API_KEY=sk-...

Local providers (lm-studio-default, gemma-reader, embeddings) need no key — just a running LM Studio instance. agent-default needs whatever its target needs: OPENAI_API_KEY as shipped.

Providers without a bundled config

Not every supported provider ships a seed config — there is nothing to seed but a model name and a key. Create one in the Admin UI (the base URL fills itself in, the model list loads once the key is there) or drop a JSON file next to the others:

llm-configs/grok.json
{
  "id": "00000001-0000-0000-0000-000000000020",
  "name": "grok",
  "provider": "XAI",
  "model": "${XAI_MODEL:grok-4}",
  "apiKey": "${XAI_API_KEY}",
  "defaultTemperature": 0.7,
  "maxOutputTokens": 8192,
  "contextWindowTokens": 256000
}

The same shape works for MOONSHOT (Kimi — set baseUrl to https://api.moonshot.cn for the mainland-China endpoint), DEEPSEEK, MISTRAL, GROQ and the rest: leave baseUrl out and the provider's own endpoint applies.

Adding your own

Drop a new JSON file into initial-data/llm-configs/, give it a unique name, and reference that name from an agent's llmConfigName. In the Admin UI you can also create and edit configs from the LLM Configs screen.

Supported providers​

Anatomy of a config​

Setting API keys​

Providers without a bundled config​

Adding your own​

Supported providers

Anatomy of a config

Setting API keys

Providers without a bundled config

Adding your own