LLM Providers
Configuring which model powers the agent — providers, keys, and exactly what the model sees
MinusX brings its own agent; the model is configured in the app. By default every workspace routes through the managed MinusX provider — add your own provider key to send LLM traffic directly from your deployment to that provider instead.
Configuring providers (Settings → Models)
The setup wizard's AI Models step, or Settings → Models, is the single place model config lives (there are no LLM environment variables):
- add one or more providers — MinusX (managed), Anthropic, OpenAI, Google, Amazon Bedrock, ~30 other registry providers, or any custom OpenAI-compatible endpoint — each with its API key,
- test the connection before saving,
- assign a model to each use case — Analyst (the main agent) and Micro tasks (titles/summaries).
Keys are stored server-side in the workspace secrets store — the config document and the browser only ever see @SECRETS/… references, never the key.
Resolution order per use case:
- The use case's explicit model assignment
- A configured MinusX provider — fully managed, no assignments needed (the gateway routes models and prompts per use case)
- The MinusX gateway default — an unconfigured workspace routes there and gets a clear authentication error pointing back at Settings → Models
For Amazon Bedrock, authenticate with a Bedrock API key (bearer token) plus the AWS region.
The heavier analyst model drives answer quality far more than any other setting. Use a strong frontier model for the Analyst assignment; Micro tasks can run a small, cheap model.
Supported providers
The provider/model pickers are backed by the pi-ai registry, overlaid live with the open models.dev catalog — newly released models appear in the pickers (and run) without waiting for an app update. Registry providers include anthropic, openai, google, google-vertex, amazon-bedrock, azure-openai-responses, openrouter, groq, mistral, deepseek, xai, cerebras, fireworks, and more.
OpenRouter is the widest single door — one key routes to hundreds of models across vendors.
The recommended starting points (also offered by the install.sh interview — this table renders from the same compatibility.json the script uses):
| Provider | Models | Default (Lite / Core / Advanced) |
|---|---|---|
minusx — Fully managed — one API key, model choice and routing handled by the MinusX gateway. | managed by the gateway | — |
anthropic | claude-haiku-4-5, claude-sonnet-5, claude-opus-4-8 | claude-haiku-4-5 / claude-sonnet-5 / claude-opus-4-8 |
openai | gpt-5.6-luna, gpt-5.6-terra, gpt-5.6-sol | gpt-5.6-luna / gpt-5.6-terra / gpt-5.6-sol |
google | gemini-3.5-flash, gemini-3.1-pro-preview | gemini-3.5-flash / gemini-3.1-pro-preview / gemini-3.1-pro-preview |
amazon-bedrock | global.anthropic.claude-haiku-4-5-20251001-v1:0, global.anthropic.claude-sonnet-5, global.anthropic.claude-opus-4-8 | global.anthropic.claude-haiku-4-5-20251001-v1:0 / global.anthropic.claude-sonnet-5 / global.anthropic.claude-opus-4-8 |
custom — Self-managed OpenAI-compatible endpoint (Ollama, vLLM, LiteLLM, …). | any model your endpoint serves | — |
Local & custom OpenAI-compatible endpoints
Any OpenAI-compatible server — Ollama, vLLM, llama.cpp, LM Studio, LiteLLM, or an internal gateway — can power the agent: add a provider with type Custom, set its base URL (e.g. http://localhost:11434/v1) and optional API key, then assign its model id (e.g. qwen3:32b) to a use case. Advanced knobs (context window, max tokens, reasoning flag, image input, extra headers, OpenAI-compat quirks) are available on the assignment's custom-model options.
The agent is tool-heavy: answer quality tracks the model's tool-calling ability. Large frontier models work best; small local models will noticeably degrade — local support is for when data-locality requirements demand it, not a free lunch. Point Micro tasks at a small local model freely; be more careful with the Analyst.
What the model sees
Requests to your configured provider contain:
- Schema metadata for connections the active context whitelists (table/column names, column statistics) — not row data, except query results the agent explicitly reads while answering
- Knowledge Base content you've written (definitions, quirks, documentation)
- Conversation content — messages, and query results returned by the agent's tools
- Chart images rendered for the conversation, where applicable
Your warehouse credentials are never part of an LLM request. With your own provider key, MinusX doesn't sit in the path and cannot see this traffic; on the managed MinusX provider, traffic routes through the MinusX gateway.