LLM config JSON

The format of a file in initial-data/llm-configs/. One file per config.

Example

llm-configs/claude-default.json
{
  "id": "00000001-0000-0000-0000-000000000002",
  "name": "claude-default",
  "provider": "ANTHROPIC",
  "model": "${CLAUDE_MODEL:claude-sonnet-4-6}",
  "baseUrl": "https://api.anthropic.com",
  "apiKey": "${ANTHROPIC_API_KEY}",
  "defaultTemperature": 0.7,
  "maxOutputTokens": 8192,
  "additionalParams": {
    "thinking": "adaptive",
    "effort": "high"
  },
  "contextWindowTokens": 200000,
  "capabilities": ["TOOL_CALLING", "VISION", "DOCUMENTS"],
  "rateLimit": {
    "maxConcurrentRequests": 3
  },
  "retry": {
    "enabled": true,
    "maxAttempts": 5,
    "baseBackoffMillis": 2000,
    "maxBackoffMillis": 30000
  },
  "fallbackModels": ["openai-default", "gemini-default"]
}

A minimal local config (no key needed):

llm-configs/lm-studio-default.json
{
  "id": "00000001-0000-0000-0000-000000000001",
  "name": "lm-studio-default",
  "provider": "LM_STUDIO",
  "model": "openai/gpt-oss-120b",
  "baseUrl": "http://localhost:1234",
  "apiKey": "lm-studio",
  "defaultTemperature": 0.7,
  "maxOutputTokens": 8192,
  "additionalParams": {},
  "contextWindowTokens": 131072
}

An alias carries no provider settings of its own and just names another config. The bundled agent-default, which every shipped agent references, is one:

llm-configs/agent-default.json
{
  "id": "00000001-0000-0000-0000-000000000000",
  "name": "agent-default",
  "isAlias": true,
  "delegatesTo": "openai-default"
}

Fields

Field	Type	Description
`id`	UUID	Stable identifier. Use a fresh UUID for a new config.
`name`	string	Unique name; what an agent references via `llmConfigName`.
`provider`	enum	One of `ANTHROPIC`, `OPENAI`, `AZURE_OPENAI`, `GOOGLE_GEMINI`, `LM_STUDIO`, `GROQ`, `OLLAMA`, `MISTRAL`, `DEEPSEEK`, `XAI` (Grok), `MOONSHOT` (Kimi), `TOGETHER`, `OPENROUTER`, `PERPLEXITY`, `FIREWORKS`.
`model`	string	Model id. Supports `${VAR}` / `${VAR:default}`.
`baseUrl`	string	API endpoint. Optional — left out, the provider's own endpoint applies (`https://api.openai.com`, `https://api.mistral.ai`, `https://api.groq.com/openai`, `http://localhost:11434` …). Set it for a proxy, a local server, or an Azure resource, which has no default.
`apiKey`	string	API key — almost always an env-var placeholder like `${ANTHROPIC_API_KEY}`. Literal keys are stored encrypted in the Admin UI (the CLI stores them as-is — use placeholders there).
`type`	enum?	`CHAT` (default), `EMBEDDING` or `SPEECH_TO_TEXT`. Embedding configs power the vector store; speech-to-text configs turn a recording into text through `LlmTranscription` (since 0.5.3). Only chat configs have sampling settings.
`capabilities`	string[]?	What the model reads and does: any of `TOOL_CALLING`, `VISION` (images), `DOCUMENTS` (PDF), `AUDIO_INPUT`. `VISION` and `DOCUMENTS` decide whether an image or PDF sent with a message reaches the model as content or as a placeholder line. Omitted means not declared — the provider's default applies (Anthropic and OpenAI: tool calling, vision, documents; Gemini: those plus audio; Azure OpenAI: tool calling, vision; local servers and routers: tool calling only). A declared list, `[]` included, always wins over that default. Chat configs only.
`defaultTemperature`	number	Sampling temperature (e.g. `0.7`).
`maxOutputTokens`	int	Max tokens the model may generate per response.
`contextWindowTokens`	int	Token budget used to size the working-memory window.
`additionalParams`	object	Provider-specific options. For Anthropic `thinking` (`adaptive`/`disabled`) and `effort` (`low`, `medium`, `high`, `xhigh`, `max`); for OpenAI also `reasoning_summary` (`auto`, `concise`, `detailed`, `none`); for every OpenAI-compatible provider `reasoning_effort` (`none`, `minimal`, `low`, `medium`, `high`, `xhigh` — which levels apply depends on the model; local servers ignore it, and OpenAI is only sent it with a reasoning model, since it rejects it elsewhere).
`rateLimit`	object?	Optional. `maxConcurrentRequests` caps in-flight calls.
`retry`	object?	Optional. `enabled`, `maxAttempts`, `baseBackoffMillis`, `maxBackoffMillis`. Omitting the block means no retry at all — the built-in defaults only apply when the block is present and enabled.
`fallbackModels`	string[]?	Optional. Other configs, by name, to try in order when this one is rate-limited (HTTP 429) or the provider is overloaded (529) — after `retry` is exhausted. Point them at a different provider: its limit is a different limit. A fallback may name an alias. Omitted or `[]` means the call fails with the rate-limit error. Chat configs only.
`isAlias` / `delegatesTo`	bool / string	Alias configs forward to another config by name — useful to give a stable name (`agent-default`) that you can repoint.

Env-var placeholders

name, model, baseUrl and apiKey support ${VAR} / ${VAR:default}, resolved at use time — the raw placeholder stays in storage. See how API keys resolve and the environment variables page.

See also the LLM configuration reference for the bundled configs and supported providers.

Example​

Fields​

Example

Fields