Models and catalog behavior

Every configured provider must authenticate and complete live discovery. RouteKit turns each native model into a canonical namespaced ID such as openai/gpt-5.5 or openrouter/anthropic/claude-sonnet-4.5.

An optional top-level modelPolicy filters those canonical IDs. Its anchored * glob can span /, so openrouter/moonshotai/* matches nested OpenRouter model IDs. The inclusive allowlist defaults to all models when omitted or empty; the denylist removes matches and has precedence.

Only effective routable models appear in /v1/models, routekit models list, provider status, native client pickers, aliases, defaults, and request routing. Policy never skips provider authentication or discovery. If policy removes all discovered models, startup fails instead of creating an empty configured catalog.

Claude Code model selection

routekit claude install derives Claude's picker entries from this same policy-filtered catalog. The entries use the reversible anthropic.routekit.<canonical-routekit-id> custom-model spelling required by Claude Code; no synthetic claude-* aliases or effort-specific models are created. Direct Claude invocations may instead use a bare native ID when exactly one RouteKit route owns it, for example claude --model gpt-5.6-sol. If it is ambiguous, Claude receives an error naming the canonical choices.

Codex model selection

Codex uses Responses-compatible routes only. An explicit namespaced model stays exact: RouteKit either uses that advertised route or rejects it.

When a Codex launch omits the model, RouteKit first honors a compatible configured defaultModel. Otherwise it considers only models that advertise text output, tool support, and a compatible Responses route. It stays on the configured provider and uses provider-authored preference rank when available, then advertised model creation time. Only when that provider has no compatible candidate can selection widen to another provider in the same billing class. It never crosses from a subscription route to a metered API route.

Direct OpenAI discovery does not expose operation capabilities. Only when an implicit Codex launch needs missing classification or recency data, RouteKit may request OpenRouter's public model metadata with Accept: application/json. It sends no API key, RouteKit token, configured model ID, prompt, or account data. Explicit selection and a compatible configured default skip this request.

Reasoning support by route

Provider reasoning data is not universally portable. When a conversation moves between Claude, chat-based providers, and native Responses providers, RouteKit keeps the portable messages and tool history. Opaque encrypted reasoning is forwarded only to the same originating provider and native model; it is omitted from an incompatible route instead of causing the conversation to fail.

Reasoning effort values come from discovered or explicitly configured model capabilities. Unsupported efforts fail before provider routing. Claude Code uses its native effort selector, and Codex uses its native picker levels.

Inspect available models

routekit models list
routekit models list --provider openai
routekit models info <provider/model>

models info reports the canonical and native IDs, provider, account class, billing mode, capabilities, and reasoning metadata without returning credentials.