How to choose models

A model is written provider:model. By default every phase uses anthropic:claude-opus-5-5. The model must support tool calling.

Use one model for everything

Set STROMBOLI_MODEL and the provider’s key:

STROMBOLI_MODEL=openai:gpt-5
OPENAI_API_KEY=sk-...

Use a fast model for triage, a thorough one for the root cause

An investigation has two phases: triage, until the mitigation is published, where speed matters most, and investigate, for the root cause and the report. Give each its own model and effort:

STROMBOLI_TRIAGE_MODEL=anthropic:claude-haiku-5-5
STROMBOLI_TRIAGE_EFFORT=low
STROMBOLI_INVESTIGATE_MODEL=openai:gpt-5
STROMBOLI_INVESTIGATE_EFFORT=high

The investigate model takes over the triage conversation, so it sees everything triage saw. Unset values fall back to STROMBOLI_MODEL and STROMBOLI_EFFORT. Effort is low, medium, high, xhigh or max; leave it unset for the provider’s default.

Use every model with one key (Vercel AI Gateway)

An id with a slash and no provider: prefix goes through Vercel AI Gateway, as in the AI SDK:

AI_GATEWAY_API_KEY=...
STROMBOLI_TRIAGE_MODEL=anthropic/claude-haiku-5.5
STROMBOLI_INVESTIGATE_MODEL=openai/gpt-5

Use a specific provider

Provider Model id Set
Anthropic anthropic:claude-opus-5-5, or a bare claude-… id ANTHROPIC_API_KEY
OpenAI openai:gpt-5 OPENAI_API_KEY; OPENAI_BASE_URL for Azure or a proxy
Google Gemini google:gemini-2.5-pro GEMINI_API_KEY
OpenRouter openrouter:meta-llama/llama-3.3-70b-instruct OPENROUTER_API_KEY
DeepSeek deepseek:deepseek-v4-pro DEEPSEEK_API_KEY; optional DEEPSEEK_BASE_URL
Vercel AI Gateway openai/gpt-5 or gateway:<id> AI_GATEWAY_API_KEY or VERCEL_OIDC_TOKEN; optional AI_GATEWAY_BASE_URL
Ollama ollama:qwen3:32b OLLAMA_BASE_URL (default http://localhost:11434/v1)

Use any OpenAI-compatible API

Pick a provider name, and set its base URL (and key, if it needs one). The name is upper-cased, with - written _:

STROMBOLI_PROVIDER_GROQ_BASE_URL=https://api.groq.com/openai/v1
STROMBOLI_PROVIDER_GROQ_API_KEY=gsk_...
STROMBOLI_MODEL=groq:llama-3.3-70b-versatile

Check it

On startup the server logs the models it resolved:

level=INFO msg="stromboli listening" addr=:8080 triage_model=anthropic:claude-haiku-5-5 investigate_model=openai:gpt-5

If a model cannot be resolved (unknown provider, missing key), the server still starts, logs LLM is not configured: briefs work, investigations are disabled, and starting an investigation fails with no LLM is configured.

Each llm step of an investigation (stromboli show <id> --json) records the task, model and token counts of that turn.

See also

Edit this page on GitHub ↗