# LLM providers & model discovery

> Phase 9 W10 · design: [design-phase-9-w10-llm-providers.md](design-phase-9-w10-llm-providers.md)

An **LlmProvider** is a reusable LLM *route*: `{ name, api kind, baseUrl?,
credentialName, scope }`. The API key is never stored on the provider — it
lives in the encrypted [credential](credentials.md) store, referenced by
name, and must be distributable (it gets pushed to machines).

## API kinds vs. agent support

Three kinds — `openai-chat`, `openai-responses`, `anthropic` — because the
agent tools differ: codex speaks Responses-only, Claude Code needs the
Anthropic flavor, etc. Each agent target accepts only its supported subset
(the machine page filters automatically).

## Model discovery

`获取模型 / fetch models` (pre-save or per-provider) resolves the credential,
fetches the endpoint's model list, and returns sorted ids/display names.
Failures surface as `502/504` with a clear code. This is the platform's
second deliberate outbound HTTP surface (the skill marketplace is the first):
GET-only, http(s) only, time-boxed, body-capped.

## Multi-model

A machine's runtime config takes a default model plus optional extra model
ids. What each agent does with the set differs (dsh: a live per-provider
catalog; claude-code: an allowlist; others: single model) — details in
[Harness runtime management](harness-runtime-management.md), and the in-chat
model picker lists only platform-configured models.