The model-provider docs at a glance: Model Providers — the architectural pattern · Provider Configuration — framework config & chat-client factories · Model Provider Setup — operational setup & troubleshooting · Model Provider Settings — the settings UI. This page: the settings UI.

AI Model Provider Settings

Current state first. The Settings → Models tab this page was written to specify (ModelsSettingsTab) no longer exists. Today a user wires AI in three places: provider keys in the Providers app's ProviderSetup view, models in the /Provider/AiModels catalog, and a CLI harness login through /login in the chat composer. What is wired below names the code. Everything from Two provider kinds through the Scope note is the original design, kept as history: its card layouts, diagrams and state machine describe a UI that is not served. The AI engine code named on this page lives in MeshWeaver.Plugins, not in this repository.

This page began as the implementation spec for a single Settings → Models destination: adding API keys, enabling specific models, and connecting CLI-based providers like Claude Code and GitHub Copilot.

Setting up models (admin or user)? This page is the UI design spec. For the operational how-to — provider/model mesh nodes, the system/space/user layers, which query goes where in a user's namespace, the open-weight tier choices, and the install-time config gaps — read Setting Up Model Providers.

Two provider kinds, two different UIs (original design — historical)

The fundamental insight driving this design: API providers and CLI providers need completely different layouts. Rendering both as a key/endpoint form is wrong.

Two provider kinds — API providers add a key and pick models; CLI providers delegate to their own auth flow.

What is wired

What is served today. The original design shipped and was later replaced; these bullets describe the code as it stands on MeshWeaver.Plugins main.

• The catalog chain registers all providers in MemexConfiguration via .AddAnthropic().AddAzureFoundry().AddAzureOpenAI().AddOpenAI().AddClaudeCode().AddCopilot(), each gated by its Features:Ai:Providers:* / Features:Ai:Clis:* flag.
• Where a user enters provider keys today is the Providers app's ProviderSetup view (MeshWeaver.Plugins Providers/ProvidersApp/Source/ProviderSetupAreas.cs; its rules live in the pure ProviderSetup.cs, tested by Providers/ProvidersApp/Test/ProviderSetupTests.cs). The AI menu's Models entry links the scope-tabbed model catalog at /Provider/AiModels (AiCatalogLayoutAreas.ModelsArea). The settings tab this page originally specified, ModelsSettingsTab with its BuildCliCard, no longer exists in either repository.
• A CLI harness login is no longer a settings card: it is the /login harness command in the chat composer (ThreadChatView.StartHarnessConnect in MeshWeaver.Plugins src/MeshWeaver.Blazor.Chat/), which drives the same ConnectSessionManager.StartConnect backend described below.
• Per-user credentials live in ModelProvider mesh nodes (ModelProviderService, ModelProviderNodeType), with keys encrypted via Ai:KeyProtection:MasterKey (sourced from Key Vault).
• The CLI connect flow is implemented: ConnectSessionManager drives ClaudeConnectStrategy / CopilotConnectStrategy (src/MeshWeaver.AI/Connect/).

Design

1 The ProviderKind seam

An explicit ProviderKind enum (Api | Cli, declared in BuiltInLanguageModelProvider.cs) sits on the provider catalog entry, replacing the implicit "has a key form" test. CLI providers (AddClaudeCode, AddCopilot) report Cli; everything else reports Api. The settings card that originally switched on ProviderKind (ModelsSettingsTab) has been removed; the enum remains on the catalog entry (MeshWeaver.Plugins src/MeshWeaver.AI/BuiltInLanguageModelProvider.cs).

2 API providers — key/endpoint form + model list

• Render the endpoint/key form. Azure providers also take an endpoint; Anthropic and OpenAI take only a key.
• With a key entered, Fetch models calls ProviderModelLister.ListModels(endpoint, apiKey, providerName) — an IObservable<IReadOnlyList<string>> — and shows the returned ids as checkable rows so the user enables the ones they want. The selection is persisted on the ModelProvider node by Save provider.
• API providers are the only kind that produce a model list — CLI providers deliberately do not.

3 CLI providers — login status + delegate to the CLI

No key form. No model list. The card shows exactly two states:

• Logged in → green "Connected as …" indicator + a "Re-connect / Log out" affordance.
• Not logged in → a "Connect / Log in" button that drives the CLI's native auth flow.

The backend lives in src/MeshWeaver.AI/Connect/:

IConnectStrategy — one implementation per CLI.

• ClaudeConnectStrategy: spawns claude under the user's CLAUDE_CONFIG_DIR, probes the CLI (or its .credentials.json) for login status, and if absent runs claude setup-token (or /login), scrapes the auth URL, surfaces it, and captures the pasted code/token.
• CopilotConnectStrategy: runs the Copilot SDK device-flow — surfaces the device URL + code, polls GetAuthStatusAsync.
• Both reuse the subprocess shape from MeshPlugin/KernelExecutor (RedirectStandardInput). The process boundary runs on the Process IIoPool — InvokeBlocking for the spawn, Invoke for the stdin write — carrying only the async leaves, with the auth-URL/token scrape composed as an ordinary observable. Never Observable.FromAsync, which is forbidden outside IoPool.

ConnectSessionManager — a mesh-scoped singleton that holds the live Process between "show URL" and "paste code", keyed per user (instance ConcurrentDictionary, never static), with a 5-minute timeout that calls Kill(entireProcessTree:true).

Login-status check is the cheap, always-on part: each card calls strategy.IsLoggedIn(userConfigDir) on render. Only the not-logged-in branch shows the login button.

On token capture → ModelProviderService.CreateProvider(ownerPath, "ClaudeCode"|"Copilot", token) (already calls Protect() on the key); re-connect uses RotateKey. The CLI agent factory already injects this.

4 Fix the missing icon

The AI settings tab is registered without an Icon. Add one where the Settings tabs are declared (the AI/Models tab registration in the portal settings) — a FluentUI Sparkle or Bot icon, consistent with other tabs.

UI — inline login

The CLI login expands inside the provider card — no modal, no side panel. This is the lightest-weight option and keeps the user in context. The card is a small state machine.

Models tab layout

Settings ▸ ✦ AI / Models

API providers — add a key, choose models
┌─ Azure AI Foundry ──────────────────────────────── [API] ┐
│ Endpoint  https://….services.ai.azure.com   [Save]  ✓     │
│ API key   ••••••••••••••••                                 │
│ Models    ☑ gpt-4o   ☑ o3-mini   ☐ embed-v-4-0             │
└────────────────────────────────────────────────────────────┘
┌─ Anthropic ─────────────────────────────────────── [API] ┐
│ API key  •••••••• [Save]    Models  ☑ claude-opus-4         │
└────────────────────────────────────────────────────────────┘

CLI providers — log in with your subscription (no key, no model list)
┌─ Claude Code ───────────────────────────────────── [CLI] ┐
│ ● Not connected — uses your Claude subscription            │
│                                  [ Connect Claude Code ]   │
└────────────────────────────────────────────────────────────┘
┌─ GitHub Copilot ────────────────────────────────── [CLI] ┐
│ ✓ Connected as @rbuergi                    [ Disconnect ]  │
└────────────────────────────────────────────────────────────┘

CLI card — connecting states

Claude Code (paste-a-code flow):

┌─ Claude Code ──────────────── [CLI] ┐
│ ● Connecting…                       │
│ 1  Authorize in your browser:       │
│    claude.ai/oauth/auth?…  [Copy][Open]
│ 2  Paste the code Claude shows:     │
│    [ __________________ ] [Submit]  │
│    ⏳ waiting for code… (4:58)       │
└──────────────────────────────────────┘

GitHub Copilot (device code, auto-poll — nothing to paste):

┌─ GitHub Copilot ───────────── [CLI] ┐
│ ● Connecting…  enter code at         │
│   github.com/login/device            │
│        ┌───────────────┐             │
│        │  AB12-CD34     │  [Copy]     │
│        └───────────────┘             │
│   ⏳ auto-checking…                   │
└──────────────────────────────────────┘

Connected: ✓ Connected as <name> [ Disconnect ] · Error/Expired: red status line + [ Retry ]

Inline state machine

NotConnected ──[Connect]──▶ Connecting ──(code submitted / device poll OK)──▶ Connected
     ▲                          │  ▲                                              │
     └──────[Disconnect]────────┘  └──(5-min timeout · Cancel · auth error)──▶ Error/Expired
                                            └──────────────[Retry]──────────────────┘

• Connecting is driven by ConnectSessionManager + the provider's IConnectStrategy. It holds the live CLI Process, exposes the auth URL (and, for Copilot, the device code), and either accepts a pasted code (RequiresPastedCode = true) or polls auth status. A 5-minute timeout disposes the process via Kill(entireProcessTree:true).
• On success the strategy returns the token → ModelProviderService.CreateProvider/RotateKey (encrypted with the Key-Vault master key) → the card flips to Connected reactively.
• The card body is chosen off strategy.RequiresPastedCode (paste field vs. device code) and IsLoggedIn(userConfigDir) (Connected vs. NotConnected on first render).

Where the code lives

Testing

No mocks. Use MonolithMeshTestBase / AITestBase.

Three test scenarios:

1. Rendering — the ProviderSetup rules (ProviderSetupTests in MeshWeaver.Plugins Providers/ProvidersApp/Test/) decide which fields and models a provider card shows. Assert on what the pure helper returns.

2. Connect flow — a committed fake CLI (prints an auth URL, reads stdin, prints a token) drives IConnectStrategy: IsLoggedIn returns false → connect → strategy captures the token → a ModelProvider node is written with an enc:-tagged key that round-trips through ChatClientCredentialResolver.

3. Login status — with the fake CLI reporting "logged in", the card renders the connected state and shows no login button. Real-CLI end-to-end is gated by CLAUDE_CONNECT_E2E=1 (developer-run only).

Scope note

(Historical.) What originally shipped was Phase 1: per-user CLI Connect plus the Models-tab rework, meaning the UI and the CLI login backend. The Models tab has since been removed; the CLI login backend remains and is driven from /login. The ProviderKind layout split was the quick visible win; the CLI login backend (ConnectSessionManager + strategies) was the substantive part.

Model picker: provider-first selection and empty state

Providers and models are mesh nodes discovered via a nodeType: fan-out query — not a flat config list. The picker lists providers first; selecting one loads only that provider's models. When nothing is configured it routes the user directly to Settings.

Providers and models are nodes

• A provider is a ModelProvider node; its models are child LanguageModel nodes nested beneath it.
• Canonical path: the platform catalog at Provider/{Provider}/{modelId} (e.g. Provider/AzureFoundry/gpt-5), a space provider at {space}/Provider/{Provider}/{modelId} (e.g. Systemorph/Provider/AzureFoundry/gpt-5), and a user's own at {user}/_Memex/{Provider}/{modelId} (e.g. rbuergi/_Memex/ClaudeCode).
• The built-in system catalog at Provider/{provider}/{model} is a DB-synced NodeType catalog: BuiltInLanguageModelProvider is the sync source, ModelStaticRepoSource imports it into the Provider partition on boot, and the DB serves it thereafter, so configuration values (AzureFoundry:Models) materialise as real, queryable MeshNodes — there is no parallel, non-mesh path.

Provider-first, lazy model load

The picker does not eager-load every model from every space. It operates in two steps:

1. List providers — fan-out nodeType:ModelProvider scope:descendants over the Provider catalog and every space the user can read. Listing providers (not models) is cheap, making it safe to broaden across spaces without loading the full model universe.

2. Select a provider → load its models — the provider's path is appended to {user}/_Memex/Selection.SelectedProviderPaths; the selected-path query namespace:{providerPath} nodeType:LanguageModel scope:selfAndDescendants (AgentPickerProjection) loads just that provider's models.

Selection is the per-user selection store at {user}/_Memex/Selection. It is seeded empty at onboarding, so the RoutingGrain NotFound: {user}/_Memex/Selection read no longer occurs against a missing node.

Empty state → Settings

When the provider fan-out returns nothing (no provider configured), the model picker does not render an empty dropdown. Instead it shows an actionable empty state: "No model provider configured" with a link that navigates to Settings → Models (action://settings/models / the OnActionLink hook) where the user can add an API key, connect a CLI, or select an org provider.

Org-default provider

An admin may pre-create an org provider node — <org>/Provider/AzureFoundry with model children sourced from your Azure AI Foundry resource (endpoint https://<foundry-account>.services.ai.azure.com/models, key in Key Vault) — that every user with read access can select. This complements per-user BYO-key and Connect flows; it does not replace the empty-state link. ModelProvider is a creatable node type (search-hidden), so it can be authored in the UI by anyone with Permission.Api, not only through configuration.

Managing a provider's models

Selecting a provider in Settings → Models lists its child LanguageModel nodes, where the user can add, remove, or enable individual models (CRUD on {provider}/{modelId} nodes).

• Fetch from provider — where the provider exposes a list API (Azure Foundry / Azure OpenAI deployment list, OpenAI / Anthropic list-models), a "Fetch models" action queries it and offers the returned IDs for import as child LanguageModel nodes, so the catalog does not need to be hand-typed.
• Refresh — re-runs the fetch and reconciles against the current children (adds newly-deployed models, flags dropped ones). Manual button now; can become periodic later.