LLM Setup

Connect a model provider¶

Mewbo needs two keys to start, an API key and a default model, set in configs/app.json.

Minimum configuration¶

Start from the shipped example rather than an empty file.

cp configs/app.example.json configs/app.json

Then set these two keys in configs/app.json.

configs/app.json

{
  "llm": {
    "api_key": "sk-ant-xxxxxxxx",
    "default_model": "anthropic/claude-sonnet-4-6"
  }
}

That's it. LiteLLM routes anthropic/claude-sonnet-4-6 to the Anthropic API using the key you provide, and direct provider access needs no api_base URL.

Edit it in the console

The console's Settings → Models & Inference panel writes this same config. Its fields map one to one onto the keys documented on this page.

The Models & Inference panel of the Mewbo console Settings screen, with fields for the API base, a configured (write-only) API key, and the default, action-plan, tool, and title model IDs

Optional LLM configuration¶

Key	Purpose	Notes
`llm.api_base`	Base URL override.	Only needed when using a proxy (LiteLLM, Bifrost). Leave empty for direct provider access.
`llm.action_plan_model`	Model for plan generation.	Falls back to `llm.default_model` if unset.
`llm.tool_model`	Model for tool execution.	Falls back to `llm.action_plan_model`, then `llm.default_model`.
`llm.reasoning_effort`	Default reasoning effort level.	Values: `low`, `medium`, `high`, `none`.
`llm.reasoning_effort_models`	Allowlist for reasoning effort.	Supports exact matches and `*` suffix wildcards.
`llm.proxy_model_prefix`	Prefix prepended to model names when routing through a proxy.	Default: `"openai"`. Set to match your proxy's expected provider prefix. Falls back to `"openai"` when empty.

Model fallback¶

When the primary model fails with a retryable error, Mewbo walks an ordered fallback list before giving up. That is what carries a run through a rate limit or a provider outage. It is off by default and needs an explicit opt-in in configs/app.json.

configs/app.json

{
  "llm": {
    "default_model": "anthropic/claude-sonnet-4-6",
    "fallback": {
      "enabled": true,
      "models": [
        "openai/gpt-4o",
        "anthropic/claude-haiku-4-5"
      ]
    }
  }
}

The flat llm.fallback_models array is also honoured. It supplies the ladder when fallback.models is empty, and a non-empty value there counts as fallback enabled. Prefer the typed fallback block.

Each fallback gets one attempt, and the error class decides what happens next.

Transient errors, such as rate limits or timeouts, retry the primary model first, then cascade.
Context overflow skips straight to the next model.
Auth errors abort immediately, since retrying the same provider wouldn't help.

agent.llm_call_retries sets the retries against the primary before cascading, and defaults to 2.

MCP setup¶

MCP servers are optional, and each one adds external tools to the registry. Create configs/mcp.json or run /mcp init in the CLI, then start a client once to discover tools and cache the manifest under ~/.mewbo/. MCP Tools has the file schema and the per-project merge.

LiteLLM provider support¶

The LLM layer runs on LiteLLM through langchain-litellm. Model IDs use provider/model syntax, for example anthropic/claude-sonnet-4-6, openai/gpt-4o, or mistral/mistral-small. LiteLLM routes each one to the correct API. Behind a proxy with no provider prefix on the model name, llm.proxy_model_prefix is prepended to route to an OpenAI-compatible endpoint.