feat(models): enforce provider visibility policy

Author: erichare

.secrets.baseline

@@ -2253,7 +2253,7 @@
         "filename": "src/backend/tests/unit/services/variable/test_service.py",
         "hashed_secret": "30abc0a833efea23496b4d226fffa2f90c0855c0",
         "is_verified": false,
-        "line_number": 43,
+        "line_number": 48,
         "is_secret": false
       }
     ],
@@ -2451,7 +2451,7 @@
         "filename": "src/backend/tests/unit/test_unified_models.py",
         "hashed_secret": "e9a5f12a8ecbb3eb46eca5096b5c52aa5e7c9fdd",
         "is_verified": false,
-        "line_number": 527,
+        "line_number": 549,
         "is_secret": false
       }
     ],
@@ -6637,7 +6637,7 @@
         "filename": "src/lfx/src/lfx/base/models/unified_models/model_catalog.py",
         "hashed_secret": "665b1e3851eefefa3fb878654292f16597d25155",
         "is_verified": false,
-        "line_number": 469,
+        "line_number": 498,
         "is_secret": false
       }
     ],
@@ -7242,5 +7242,5 @@
       }
     ]
   },
-  "generated_at": "2026-07-16T15:57:06Z"
+  "generated_at": "2026-07-16T21:26:55Z"
 }

docs/docs/Lfx/extensions-manifest.mdx

@@ -37,12 +37,15 @@ Use the `$schema` reference in your manifest so editors can autocomplete and val
 | `name` | string | yes | Human-readable display name shown in the Langflow palette. 1-200 chars. |
 | `description` | string \| null | no | Optional one-paragraph summary, max 2000 chars. |
 | `lfx` | object | yes | [Compatibility declaration](#lfx-compatibility-declaration) against the BUNDLE_API contract. |
-| `bundles` | array | yes | [Bundle list](#bundles). v0 accepts **exactly one** bundle. |
+| `bundles` | array | no | [Bundle list](#bundles). v0 accepts at most one bundle; omit it for a provider-only extension. |
+| `providers` | array | no | [Model providers](#providers) contributed to the unified model-provider registry. |
 | `capabilities` | object | no | [Optional capability flags](#capabilities). Defaults to all-false. |
 | `$schema` | string | no | Optional pointer to this JSON Schema; editors use it for autocomplete. |
 
 `additionalProperties: false` — any field not listed here is rejected with a typed error. Reserved names (`services`, `routes`, `hooks`, `starterProjects`, `userConfig`) are documented under [Deferred fields](#deferred-fields) and surface a more specific error code.
 
+An extension must declare at least one entry in either `bundles` or `providers`. It can declare both when one package contributes components and model providers.
+
 ## `lfx`: compatibility declaration
 
 ```json
@@ -68,7 +71,66 @@ The runtime compares `str(BUNDLE_API_VERSION)` against this list. A mismatch fai
 | `name` | string | Bundle name; addressable as `ext:<bundle>:<Class>@<slot>`. Lowercase snake_case, starts with a letter, 2-64 chars. |
 | `path` | string | Path to the bundle directory, relative to the manifest. Must not start with `/` or contain `..`. |
 
-v0 enforces `minItems: 1, maxItems: 1`; multi-bundle extensions are rejected with `multi-bundle-deferred-in-this-milestone` and ship in a later epic.
+v0 enforces `maxItems: 1`; multi-bundle extensions are rejected with `multi-bundle-deferred-in-this-milestone` and ship in a later epic. The list can be empty when the extension declares at least one model provider.
+
+## `providers`
+
+A provider-only extension can add a provider without adding a component bundle:
+
+```json
+{
+  "$schema": "https://schemas.langflow.org/extension/v1.json",
+  "id": "lfx-acme-models",
+  "version": "0.1.0",
+  "name": "Acme Models",
+  "lfx": { "compat": ["1"] },
+  "providers": [
+    {
+      "name": "Acme",
+      "provider_id": "acme.models",
+      "display_name": "Acme Models",
+      "aliases": ["Acme Legacy"],
+      "metadata": {
+        "icon": "Bot",
+        "variables": [],
+        "mapping": {
+          "model_class": "ChatAcme",
+          "model_param": "model"
+        }
+      },
+      "api_key_required": false,
+      "model_class": {
+        "module": "langchain_acme",
+        "attr": "ChatAcme",
+        "install_hint": "langchain-acme"
+      },
+      "catalog_loader": "lfx_acme.catalog:load_models"
+    }
+  ]
+}
+```
+
+| Field | Type | Required | Description |
+| --- | --- | --- | --- |
+| `name` | string | yes | Canonical provider name used by existing saved flows and provider selectors. |
+| `provider_id` | string | no | Stable lowercase machine identity used by policy and extension compatibility. Must match `^[a-z0-9][a-z0-9._-]*$`. Legacy manifests can omit it and receive a deterministic ID derived from `name`. |
+| `display_name` | string | no | User-facing label. It can change without changing `provider_id` or saved-flow identity. |
+| `aliases` | array of strings | no | Unique legacy names accepted when resolving the provider's stable identity. |
+| `metadata` | object | yes | Provider icon, variables, API documentation URL, and mapping. `mapping.model_class` must be non-empty. |
+| `model_class` | object | no | Lazy LLM class import with `module`, `attr`, and optional `install_hint`. Omit it when reusing an already-registered class. |
+| `embedding` | object | no | Optional embedding class import and parameter mapping. Includes `class_name`, `module`, `attr`, `param_mapping_key`, `param_mapping`, and optional `install_hint`. |
+| `api_key_required` | bool | no | Whether unified runtime helpers reject missing API-key credentials. Defaults to `true`. |
+| `live` | bool | no | Enables always-on live model discovery. Mutually exclusive with `conditional_live`. |
+| `conditional_live` | bool | no | Enables live discovery only when the provider has a custom endpoint configured. Mutually exclusive with `live`. |
+| `live_discovery` | string | no | Lazy `module:callable` that receives `(user_id, model_type)` and returns live model rows. |
+| `validator` | string | no | Lazy `module:callable` that validates `(provider, variables, model)` and raises on invalid credentials. |
+| `catalog_loader` | string | no | Lazy `module:callable` returning the provider's static model metadata rows. |
+
+`catalog_loader` must return a flat `list[dict]`. Every row requires a non-empty `name`; `model_type` can be `llm` or `embeddings` and defaults to `llm`. Langflow overwrites any supplied `provider` value with the descriptor's canonical `name`, supplies the provider icon when omitted, and rejects duplicate `(model_type, name)` identities. Deployments that require an extension catalog can call `validate_registered_provider_catalogs()` or `get_registry_snapshot(validate_catalogs=True)` during readiness.
+
+Providers that use ambient authentication or need no credentials should set `api_key_required` to `false`. They remain eligible for unified model options when `metadata.variables` is empty or contains only optional configuration fields.
+
+Built-in providers take precedence over extension declarations. A colliding provider is skipped without preventing other providers or components in the extension from loading.
 
 ## `capabilities`
 
@@ -127,6 +189,8 @@ The loader and validator both emit typed errors keyed by the manifest field that
 | `multi-bundle-deferred-in-this-milestone` | `bundles` has more than one entry. |
 | `path-escape` | A `bundles[].path` resolves outside the manifest root (typically a symlink). |
 | `bundle-path-not-found` | `bundles[].path` does not exist or is not a directory. |
+| `provider-invalid` | A provider descriptor or one of its lazy import paths is malformed. The invalid provider is skipped. |
+| `provider-skipped` | A provider name collides with a built-in or previously loaded provider. The existing provider wins. |
 
 Run `lfx extension validate <path>` to see every error as a structured object with `code`, `message`, `location`, `hint`, and `ref_url`.
 

src/backend/base/langflow/agentic/api/router.py

@@ -10,10 +10,11 @@
 
 from fastapi import APIRouter, Depends, HTTPException, Request
 from fastapi.responses import StreamingResponse
+from lfx.base.models.provider_registry import is_api_key_optional
 from lfx.base.models.unified_models import (
     get_all_variables_for_provider,
-    get_model_provider_variable_mapping,
     get_provider_required_variable_keys,
+    get_provider_secret_variable_key,
     get_unified_models_detailed,
 )
 from lfx.log.logger import logger
@@ -47,7 +48,7 @@ class _AssistantContext:
 
     provider: str
     model_name: str
-    api_key_name: str
+    api_key_name: str | None
     session_id: str
     global_vars: dict[str, str]
     max_retries: int
@@ -63,7 +64,6 @@ async def _resolve_assistant_context(
     Raises:
         HTTPException: If provider is not configured or API key is missing.
     """
-    provider_variable_map = get_model_provider_variable_mapping()
     enabled_providers, _ = await get_enabled_providers_for_user(user_id, session)
 
     if not enabled_providers:
@@ -87,8 +87,8 @@ async def _resolve_assistant_context(
             detail=f"Provider '{provider}' is not configured. Available providers: {enabled_providers}",
         )
 
-    api_key_name = provider_variable_map.get(provider)
-    if not api_key_name:
+    api_key_name = get_provider_secret_variable_key(provider)
+    if not api_key_name and not is_api_key_optional(provider):
         raise HTTPException(status_code=400, detail=f"Unknown provider: {provider}")
 
     model_name = request.model_name or get_default_model(provider, user_id=user_id) or ""

src/backend/base/langflow/agentic/services/provider_service.py

@@ -6,12 +6,15 @@
 
 from lfx.base.models.model_metadata import CONDITIONAL_LIVE_MODEL_PROVIDERS, LIVE_MODEL_PROVIDERS
 from lfx.base.models.model_utils import get_live_models_for_provider
+from lfx.base.models.provider_registry import is_api_key_optional
 from lfx.base.models.unified_models import (
     get_model_provider_variable_mapping,
+    get_model_providers,
     get_provider_required_variable_keys,
     get_unified_models_detailed,
 )
 from lfx.log.logger import logger
+from lfx.services.model_provider_policy import ModelProviderPolicyPurpose, resolve_model_provider_policy
 from lfx.utils.secrets import secret_value_to_str
 from sqlalchemy.ext.asyncio import AsyncSession
 
@@ -61,15 +64,36 @@ async def get_enabled_providers_for_user(
     all_variable_names = {var.name for var in all_variables}
 
     provider_variable_map = get_model_provider_variable_mapping()
+    registered_providers = get_model_providers()
+    provider_candidates = [
+        *provider_variable_map,
+        *(
+            provider
+            for provider in registered_providers
+            if provider not in provider_variable_map and is_api_key_optional(provider)
+        ),
+    ]
+    provider_policy = resolve_model_provider_policy(
+        user_id=user_id,
+        providers=[*registered_providers, *provider_candidates],
+        purpose=ModelProviderPolicyPurpose.USE,
+    )
 
     enabled_providers = []
     provider_status = {}
 
-    for provider in provider_variable_map:
+    for provider in provider_candidates:
+        if not provider_policy.allows(provider):
+            continue
         # Check if ALL required variables for this provider are present
         # in either database variables or environment variables
         required_keys = get_provider_required_variable_keys(provider)
-        is_enabled = all(key in all_variable_names or os.getenv(key) for key in required_keys)
+        provider_has_variables = provider in provider_variable_map
+        is_enabled = (
+            is_api_key_optional(provider)
+            if not provider_has_variables
+            else all(key in all_variable_names or os.getenv(key) for key in required_keys)
+        )
 
         provider_status[provider] = is_enabled
         if is_enabled:

src/backend/base/langflow/api/v1/knowledge_bases.py

@@ -22,8 +22,14 @@
     get_source_class,
     registered_sources,
 )
+from lfx.base.models.provider_registry import provider_id_for
 from lfx.base.vectorstores.chroma_security import chroma_client_create_collection_kwargs
 from lfx.log import logger
+from lfx.services.model_provider_policy import (
+    ModelProviderPolicyError,
+    ModelProviderPolicyPurpose,
+    require_model_provider,
+)
 from pydantic import BaseModel, Field
 
 from langflow.api.utils import CurrentActiveUser, ingestion_run_service, knowledge_base_service
@@ -78,6 +84,41 @@
 router = APIRouter(tags=["Knowledge Bases"], prefix="/knowledge_bases", include_in_schema=False)
 
 
+def _provider_identity(provider: str) -> str:
+    """Return a stable identity for comparison without exposing registry state."""
+    normalized = provider.strip()
+    return provider_id_for(normalized) or normalized.casefold()
+
+
+def _require_create_embedding_provider(
+    request: CreateKnowledgeBaseRequest,
+    current_user: CurrentActiveUser,
+) -> str:
+    """Validate both provider representations and enforce CONFIGURE policy."""
+    flat_provider = request.embedding_provider.strip()
+    selected_provider = None
+    if request.model_selection is not None:
+        selected_provider = knowledge_base_service.get_embedding_provider(request.model_selection)
+    if not flat_provider or (
+        selected_provider is not None
+        and (
+            selected_provider == "Unknown" or _provider_identity(flat_provider) != _provider_identity(selected_provider)
+        )
+    ):
+        raise HTTPException(status_code=404, detail="Model provider not found")
+
+    try:
+        require_model_provider(
+            user_id=current_user.id,
+            provider=flat_provider,
+            purpose=ModelProviderPolicyPurpose.CONFIGURE,
+        )
+    except ModelProviderPolicyError as exc:
+        # Treat unknown, blocked, and conflicting provider identities alike.
+        raise HTTPException(status_code=404, detail="Model provider not found") from exc
+    return flat_provider
+
+
 @dataclass(frozen=True)
 class _KbGuardResult:
     """Outcome of ``_guard_kb_action`` used by routes to know the effective owner.
@@ -709,6 +750,7 @@ async def create_knowledge_base(
 ) -> KnowledgeBaseInfo:
     """Create a new knowledge base with embedding configuration."""
     try:
+        embedding_provider = _require_create_embedding_provider(request, current_user)
         kb_root_path = KBStorageHelper.get_root_path()
         kb_user = current_user.username
         kb_name = request.name.strip().replace(" ", "_")
@@ -788,7 +830,7 @@ async def create_knowledge_base(
         backend_config_value = request.backend_config or {}
         embedding_metadata = {
             "id": str(kb_id),
-            "embedding_provider": request.embedding_provider,
+            "embedding_provider": embedding_provider,
             "embedding_model": request.embedding_model,
             "model_selection": request.model_selection,
             "created_at": datetime.now(timezone.utc).isoformat(),
@@ -827,7 +869,7 @@ async def create_knowledge_base(
             # when the request didn't carry one of its own.
             persisted_selection = request.model_selection or {
                 "name": request.embedding_model,
-                "provider": request.embedding_provider,
+                "provider": embedding_provider,
             }
             await knowledge_base_service.create_record(
                 user_id=current_user.id,
@@ -857,7 +899,7 @@ async def create_knowledge_base(
             id=str(kb_id),
             dir_name=kb_name,
             name=kb_name.replace("_", " "),
-            embedding_provider=request.embedding_provider,
+            embedding_provider=embedding_provider,
             embedding_model=request.embedding_model,
             size=0,
             words=0,