Platform Provider Setup (memex)

A fresh memex deployment has no working chat until a platform AI provider is configured. Providers and their models are mesh nodes, not appsettings/Helm values — so you configure them once, through the platform, and they survive every redeploy. This guide is the operator/first-run path.

The built-in/config provider (BuiltInLanguageModelProvider) is only a sync source — it may seed default provider/model nodes on first boot, but it can be absent entirely. The source of truth is the DB, under the Provider catalog. You never have to touch config to add or change a provider.

What you are creating

The provider catalog is a NodeType catalog under the top-level Provider partition — the same shape as Agent / Skill / Harness (see NodeType Catalogs):

@Provider                                  the catalog partition (Space)
@Provider/_Policy        PartitionAccessPolicy   PublicRead = true
@Provider/Anthropic      ModelProvider           endpoint + (encrypted) key
@Provider/Anthropic/claude-opus-4-8    LanguageModel    a model, NESTED under its provider
@Provider/Anthropic/claude-sonnet-4-6  LanguageModel
@Provider/OpenRouter     ModelProvider
@Provider/OpenRouter/{slug}            LanguageModel

Models are nested inside their provider. The model the chat picker uses by default is the LanguageModel node with the lowest Order (the order:-1 convention).

Who can configure it

Platform admins (Auth:GlobalAdmins, seeded by GlobalAdminSeed) hold write on the Provider partition. Everyone else gets read via Provider/_Policy (PublicRead = true) so the model picker resolves under any user's identity. Non-admins never see or edit keys (ModelProvider.ApiKey is gated by Permission.Api).

Setup — through the mesh catalog (no config, no redeploy)

You manage providers and models in the mesh catalog (the standard search/browse UI), which supports create, edit, and a permission-gated delete (trash) affordance + keyboard shortcuts. There is no bespoke settings form.

Create the provider. Add a ModelProvider node at Provider/{ProviderName} — e.g. Provider/Anthropic. Set its Endpoint (the provider's base URL) and paste the API key (stored encrypted at rest via ProviderKeyProtector; admins with Permission.Api only).
Add models. Add one LanguageModel child per model under the provider — Provider/{ProviderName}/{modelId} (e.g. Provider/Anthropic/claude-opus-4-8). A model can be listed live from the provider (see Listing models) or added by id.
Set the default. Give the model you want the chat default Order = -1; clear any other default. The picker resolves the lowest-Order model as the default composer model.
Verify. Open chat → the /model picker lists the catalog → start a thread → it round-trips on the configured provider. Then restart the portal pod and confirm the providers, keys, and default survive — the proof that nothing is config-stamped.

Listing models from a provider

When the GUI lists a provider's available models, it calls the provider's GET {baseUrl}/models endpoint. That outbound HTTP call goes through IIoPool (never Observable.FromAsync) per Controlled IO Pooling — the same bounded, off-hub edge every external call uses.

During onboarding

A platform that has never had a provider configured prompts for one as an onboarding step: the operator creates the first Provider/{name} node (endpoint + key) and picks a default model, then chat is live. This is the same create flow as above, surfaced once at first run so a new deployment is never left with a dead chat.

Per-deployment defaults (reference)

Deployment	Default provider (`order:-1` model)	Notes
memex / memex-cloud	OpenRouter	`https://openrouter.ai/api/v1`, models auto-listed
customer portals	per customer — e.g. Anthropic (Claude via Azure Foundry)	endpoint `…/anthropic/…`, Foundry key

These are mesh nodes, set per deployment through the catalog — not baked into any config. A user can additionally configure their own providers/models in their dotfile namespace {user}/_Memex/…; their selection of which providers are active is stored at {user}/_Memex/Selection.