Add NVIDIA NIM as a provider (OpenAI-compatible)

coworker/providers/matrix.py

@@ -76,6 +76,7 @@ class ModelEntry:
     "deepseek:deepseek-v4-pro": ModelEntry("DeepSeek V4 Pro · DeepSeek"),
     "kimi:kimi-k2.6": ModelEntry("Kimi K2.6 · Moonshot"),
     "minimax:MiniMax-M2.5": ModelEntry("MiniMax M2.5 · MiniMax"),
+    "nvidia:z-ai/glm-5.2": ModelEntry("GLM-5.2 · NVIDIA NIM"),
     "qwen:qwen3-max": ModelEntry("Qwen3 Max · Alibaba"),
     "xai:grok-4.3": ModelEntry("Grok 4.3 · xAI"),
     "mistral:mistral-large-latest": ModelEntry("Mistral Large · Mistral"),

coworker/providers/registry.py

@@ -133,6 +133,17 @@ def _build_ollama(profile: dict[str, Any], secrets: Any) -> ProviderClient:
     return OpenAIProvider(api_key="ollama", base_url=base_url)
 
 
+def _build_nvidia(profile: dict[str, Any], secrets: Any) -> ProviderClient:
+    # NVIDIA NIM is OpenAI-compatible, but it uses its own key namespace and endpoint.
+    api_key = ((profile or {}).get("api_key") or "").strip() or (
+        os.environ.get("NVIDIA_API_KEY", "").strip()
+    )
+    if not api_key:
+        raise RuntimeError("No NVIDIA API key configured — add it in Settings ▸ Models.")
+    base_url = ((profile or {}).get("base_url") or "").strip() or "https://integrate.api.nvidia.com/v1"
+    return OpenAIProvider(api_key=api_key, base_url=base_url)
+
+
 def _openai_compat(vendor: str, default_base_url: str, env_key: Optional[str] = None):
     """Builder factory for vendors reached through their OpenAI-compatible API (Z AI, DeepSeek,
     Kimi, MiniMax, Qwen, xAI, Mistral). The key is resolved from the vendor's OWN profile (or its
@@ -307,6 +318,31 @@ def _compat(
         recommended_model="mistral-large-latest",
         env_key="MISTRAL_API_KEY",
     ),
+    ProviderDescriptor(
+        name="nvidia",
+        title="NVIDIA NIM",
+        needs_key=True,
+        fields=[
+            ProviderField(
+                "api_key",
+                "NVIDIA API key",
+                secret=True,
+                placeholder="nvapi-…",
+            ),
+            ProviderField(
+                "base_url",
+                "Endpoint",
+                required=False,
+                default="https://integrate.api.nvidia.com/v1",
+                placeholder="https://integrate.api.nvidia.com/v1",
+                help="Prefilled with NVIDIA's OpenAI-compatible NIM endpoint.",
+            ),
+        ],
+        build=_build_nvidia,
+        recommended_model="z-ai/glm-5.2",
+        env_key="NVIDIA_API_KEY",
+        blurb="Uses NVIDIA NIM's OpenAI-compatible API — the endpoint is prefilled, just add your key.",
+    ),
     # Resellers: many labs' models behind one key, using THEIR model namespaces (the curated
     # ids + display labels live in providers/matrix.py). TODO: add Groq and OpenRouter here
     # (+ their matrix rows) once the current provider surface is tested — deliberately

surfaces/gui/e2e/fixtures.ts

@@ -45,6 +45,7 @@ const SETTINGS = {
   model_labels: {
     "anthropic:claude-opus-4-8": "Claude Opus 4.8 · Anthropic",
     "zai:glm-5.2": "GLM-5.2 · Z AI",
+    "nvidia:z-ai/glm-5.2": "GLM-5.2 · NVIDIA NIM",
   },
 };
 
@@ -331,6 +332,8 @@ const PROVIDERS = [
   { name: "anthropic", title: "Claude (Anthropic)", needs_key: true, fields: [{ key: "api_key", label: "API key", secret: true, required: true, help: "", placeholder: "sk-…" }], configured: true, values: {}, suggested_models: ["claude-opus-4-8"], key_set_at: null, last_used_at: null },
   // zai: an OpenAI-compatible vendor — unconfigured, with a prefilled editable endpoint + blurb.
   { name: "zai", title: "Z AI (GLM)", needs_key: true, blurb: "Uses Z AI's OpenAI-compatible API — the endpoint is prefilled, just add your key.", fields: [{ key: "api_key", label: "Z AI API key", secret: true, required: true, help: "", placeholder: "" }, { key: "base_url", label: "Endpoint", secret: false, required: false, help: "Prefilled with Z AI's international endpoint.", placeholder: "https://api.z.ai/api/paas/v4", default: "https://api.z.ai/api/paas/v4" }], configured: false, values: {}, suggested_models: ["glm-5.2"], key_set_at: null, last_used_at: null },
+  // nvidia: an OpenAI-compatible vendor — unconfigured, with a prefilled editable endpoint + blurb.
+  { name: "nvidia", title: "NVIDIA NIM", needs_key: true, blurb: "Uses NVIDIA NIM's OpenAI-compatible API — the endpoint is prefilled, just add your key.", fields: [{ key: "api_key", label: "NVIDIA API key", secret: true, required: true, help: "", placeholder: "nvapi-…" }, { key: "base_url", label: "Endpoint", secret: false, required: false, help: "Prefilled with NVIDIA's OpenAI-compatible NIM endpoint.", placeholder: "https://integrate.api.nvidia.com/v1", default: "https://integrate.api.nvidia.com/v1" }], configured: false, values: {}, suggested_models: ["z-ai/glm-5.2"], key_set_at: null, last_used_at: null },
   // ollama: keyless local provider — "configured" without proving anything runs; the
   // onboarding gallery shows "No key needed" and its form is endpoint + Detect (§39).
   { name: "ollama", title: "Ollama (local models)", needs_key: false, fields: [{ key: "base_url", label: "Endpoint", secret: false, required: false, help: "", placeholder: "http://127.0.0.1:11434", default: "http://127.0.0.1:11434" }], configured: true, values: {}, suggested_models: ["qwen3-coder:30b"], key_set_at: null, last_used_at: null },

surfaces/gui/src/App.tsx

@@ -196,6 +196,7 @@ export function App() {
   // composer's "No model connected" chip. Default true so we don't flash the chip before settings
   // load; corrected by loadSettings.
   const [modelReady, setModelReady] = useState(true);
+  const [modelsLoaded, setModelsLoaded] = useState(false);
   const [surface, setSurface] = useState<
     "session" | "scheduled" | "integrations" | "audit" | "inbox" | "persona" | "settings"
   >("session");
@@ -480,9 +481,12 @@ export function App() {
         setModels(s.models || []);
         setModelLabels(s.model_labels || {});
         setModelReady(s.model_ready);
+        setModelsLoaded(true);
         if (s.surfaces) setSurfaces(s.surfaces);
       })
-      .catch(() => {});
+      .catch(() => {
+        setModelsLoaded(true);
+      });
 
   // Open Settings → Configure Models (from the composer's "No model connected" chip).
   const openModelSetup = () => openSettings("models");
@@ -1509,6 +1513,7 @@ export function App() {
               mode={mode}
               model={model}
               models={models}
+              modelsLoaded={modelsLoaded}
               modelLabels={modelLabels}
               running={running}
               connected={connected}

surfaces/gui/src/components/Composer.tsx

@@ -47,6 +47,7 @@ interface Props {
   mode: string;
   model: string;
   models?: string[];
+  modelsLoaded?: boolean;
   modelLabels?: Record<string, string>; // curated display names (raw id when absent)
   // The model is FIXED once the session has history (§17): the picker renders ONLY on a fresh
   // session; after the first turn the fact lives in the topbar subtitle (§22) — no
@@ -320,7 +321,7 @@ export function Composer(props: Props) {
     }
   };
 
-  const modelsLoaded = !!(props.models && props.models.length);
+  const modelsLoaded = props.modelsLoaded ?? !!(props.models && props.models.length);
   const modelOptions: Option[] = Array.from(
     new Set([props.model, ...(props.models || [])]),
   ).map((m) => ({

surfaces/gui/src/components/Composer.voice.test.tsx

@@ -101,4 +101,11 @@ describe("Composer voice input (§37)", () => {
     expect((await screen.findByRole("alert")).textContent).toContain("No microphone is available.");
     expect(screen.getByLabelText("Start dictation").hasAttribute("disabled")).toBe(false);
   });
+
+  it("an empty loaded model list shows the dropdown instead of the loading chip", () => {
+    render(<Composer {...props({ models: [], modelsLoaded: true })} />);
+
+    expect(screen.queryByTestId("models-loading")).toBeNull();
+    expect(screen.getByRole("button", { name: /gpt-5.6-sol/i })).toBeTruthy();
+  });
 });

surfaces/gui/src/providers/logos.ts

@@ -18,6 +18,7 @@ import mistral from "./logos/mistral.svg";
 import qwen from "./logos/qwen.svg";
 import minimax from "./logos/minimax.svg";
 import xai from "./logos/xai.svg";
+import nvidia from "./logos/nvidia.svg";
 
 export const PROVIDER_LOGOS: Record<string, string> = {
   anthropic,
@@ -33,6 +34,7 @@ export const PROVIDER_LOGOS: Record<string, string> = {
   qwen,
   minimax,
   xai,
+  nvidia,
 };
 
 export const PROVIDER_ORDER = [
@@ -43,6 +45,7 @@ export const PROVIDER_ORDER = [
   "fireworks",
   "together",
   "zai",
+  "nvidia",
   "kimi",
   "deepseek",
   "mistral",