Provider Management

A provider is any LLM backend that Open Model Prism can forward requests to. Multiple providers can be connected simultaneously; each tenant selects which providers it has access to.

Supported Provider Types

Adding a Provider

1. Go to Providers in the sidebar
2. Click Add Provider
3. Enter a name and select the provider type
4. Enter the Base URL — do not include /v1 or /api/v1 (auto-detected)
5. Enter the API Key (leave empty for local providers like Ollama)
6. Click Test Connection — the connection check probes multiple path variants and logs results
7. Click Discover Models to auto-populate the model list

URL Detection

Open Model Prism automatically detects the correct API path by probing:

• <baseUrl>/v1/models
• <baseUrl>/api/v1/models
• <baseUrl>/models

The detected path is stored as config.options.apiPath and used for all subsequent requests.

The setup wizard and provider form both warn if the URL contains a /v1 suffix and offer a one-click auto-fix.

HTTP → HTTPS Upgrade

If a connection test succeeds over HTTP but fails over HTTPS, the form suggests switching to HTTPS. If it succeeds over HTTP only, a warning is shown — the connection works but HTTPS is recommended.

Model Discovery

After a successful connection test, click Discover Models (or the provider form triggers it automatically). The discovery process:

1. Fetches /v1/models from the provider
2. For each model: looks up the local model registry (modelRegistry.js) for tier, pricing, categories, and benchmark scores
3. Falls back to models.dev API (or offline snapshot if OFFLINE=true) for additional enrichment
4. Saves results to provider.discoveredModels[]

Discovered models can be individually edited in the Models page:

• Adjust tier and priority within tier (affects routing order)
• Override input/output pricing (USD per 1M tokens)
• Assign routing categories
• Toggle visibility (visible=false hides from tenant model lists but remains visible to admins, greyed out)

Auto-Suggest

Each model in the registry has an Auto-Suggest button (wand icon) that fills tier, pricing, and categories from the built-in model registry using fuzzy matching. The matching logic runs three passes:

1. Exact ID match
2. Registry pattern is a substring of the model ID
3. Model ID is a substring of the registry pattern

Model Registry

The local model registry (server/data/modelRegistry.js) contains metadata for 60+ models:

{
  id:            'claude-opus-4-6',
  family:        'claude',
  vendor:        'Anthropic',
  tier:          'high',
  inputPer1M:    15.00,   // USD
  outputPer1M:   75.00,   // USD
  contextWindow: 200000,
  categories:    ['reasoning_formal', 'code_generation', 'system_design'],
  patterns:      ['opus46', 'opus-4-6', 'claude-opus-4-6'],
  benchmarks: {
    intelligence: 93,
    coding:       89,
    math:         88,
    speed:        25,    // 0=slow, 100=fast
  }
}

Benchmark Scores

Benchmark scores (0-100 scale) are sourced from public benchmarks (ArtificialAnalysis, LMArena, HumanEval, MATH-500) and used by the routing engine to select the most capable model for a given task type.

Connection Check

The connection check (POST /api/prism/admin/providers/:id/check) runs a detailed diagnostic:

✓ Resolved base URL: https://api.openai.com
✓ GET /v1/models returned 200
✓ Found 47 models
✓ Test chat request succeeded (model: gpt-4o-mini)
✓ Streaming test succeeded

Each line is logged and displayed in the UI. If the check fails at any step, the log shows exactly where to help diagnose the issue.

Provider Adapters

All provider adapters extend BaseAdapter and implement:

• chat(request) — non-streaming chat completions
• stream(request, onChunk) — streaming chat completions
• embed(request) — embeddings
• listModels() — model discovery

The OpenAI-compatible adapter handles the majority of providers. Ollama has a custom adapter that translates between the Ollama format and OpenAI format. Bedrock and Azure use the OpenAI adapter with different base URL and auth patterns.