One provider profile, six coding surfaces

A shared model provider does not create a shared configuration format. Claude Code reads Anthropic environment variables, Codex accepts provider definitions, ChatGPT persists a profile, and Claude Desktop expects fixed model slots. Pointing every tool at one URL leaves the difficult part unresolved.

Launchpad makes the provider profile the stable boundary. Model discovery, endpoint resolution, and credential selection happen once. Six application surfaces then receive the narrow configuration each one understands.

A provider profile owns endpoint rules

The first implementation treated the provider URL as an origin and appended paths inside each launcher. That rule breaks as soon as a URL already ends in /v1 or includes a tenant prefix. The same configured value can produce /v1/v1/models in one adapter and lose the prefix in another.

The replacement introduces one provider profile with three inputs: provider kind, base URL, and an optional models URL. The profile validates them and derives the OpenAI base, Anthropic base, and discovery endpoints.

This follows the provider seam used by OpenCode: generic settings remain separate from provider-specific discovery behavior. A launcher does not need to know where a LiteLLM catalog lives. It asks the profile for a resolved endpoint.

Discovery follows provider kind

LiteLLM exposes richer model-group metadata outside the OpenAI API surface. A generic OpenAI-compatible provider usually exposes only the standard model list. Probing both indiscriminately turns an authentication failure into an ambiguous fallback.

The explicit models URL is for deployments where inference and discovery use different hosts or prefixes. It is not a second provider base. Launch requests still use the OpenAI or Anthropic base derived from the provider URL.

Configuration has one precedence order

Launchpad owns its environment contract instead of borrowing LiteLLM-specific variable names. This keeps the configuration accurate when the provider is not LiteLLM.

export LAUNCHPAD_PROVIDER_API_KEY=...
export LAUNCHPAD_PROVIDER_URL=https://provider.example.com/v1
export LAUNCHPAD_PROVIDER_KIND=openai-compatible

launchpad config show
launchpad launch claude

The URL, kind, and models URL use environment values first, then saved settings. The key resolves from LAUNCHPAD_PROVIDER_API_KEY first and macOS Keychain second. No LITELLM_* compatibility aliases remain.

The same profile can be saved without environment variables:

launchpad config set-provider https://provider.example.com/v1 \
  --kind openai-compatible \
  --models-url https://catalog.example.com/models

The models URL is optional. Omitting it restores provider-specific discovery.

Each surface gets a narrow adapter

The provider profile ends where tool configuration begins. Adapters receive resolved endpoints, a model ID, and a credential. They do not rediscover models or append API versions.

Model IDs pass through unchanged. The picker selects provider-model-id; Claude receives that value as its model, Codex receives it through -m, and OpenCode places it under the Launchpad provider namespace.

Credentials remain process-scoped

Saved settings contain provider kind, URL, and optional models URL. They do not contain the provider key. Terminal adapters construct a child environment from the current process, remove Launchpad’s provider variables, remove conflicting tool credentials, then add only the variables required by the selected integration.

ChatGPT and Claude Desktop need credentials after the initiating CLI process exits. Their flows persist the management key in macOS Keychain and configure a helper or local service instead of writing the key into general settings JSON.

Persistent changes require restoration

ChatGPT shares configuration with Codex under CODEX_HOME. Launchpad captures the existing root model fields, writes a delimited provider block, and records the generated model catalog before restarting the desktop app.

The restore command removes only Launchpad’s managed block, restores the captured values, removes the generated catalog, and asks before restarting ChatGPT:

launchpad launch chatgpt --restore

This is not cleanup around the implementation. It is part of the adapter contract: a persistent integration is incomplete until the previous state can be recovered.

Verify the boundary

The tests exercise provider URLs with and without /v1, path prefixes, LiteLLM fallback, direct OpenAI discovery, explicit catalog URLs, environment precedence, child credential isolation, Claude request rewriting, and ChatGPT restore.

go test -race -timeout 60s ./...
go vet ./...
npm --prefix ui run build

A smoke test can stop at picker cancellation. That proves credential resolution and live catalog discovery without starting a model request.

Takeaways