Add nvidia nim

README.md

@@ -54,7 +54,9 @@ Under the hood:
 
 Model access is yours: pick a provider, paste your key, switch anytime. Supported out of the box:
 
-**OpenAI · Anthropic · Google Gemini · Inkling (Thinking Machines) · GLM (Z.ai) · DeepSeek · Kimi (Moonshot) · Qwen · MiniMax · Mistral · Grok (xAI)** - plus open-weight models via **Together** and **Fireworks**, and fully local models via **Ollama**.
+**OpenAI · Anthropic · Google Gemini · NVIDIA NIM · Inkling (Thinking Machines) · GLM (Z.ai) · DeepSeek · Kimi (Moonshot) · Qwen · MiniMax · Mistral · Grok (xAI)** - plus open-weight models via **Together** and **Fireworks**, and fully local models via **Ollama**.
+
+NVIDIA NIM provides hosted OpenAI-compatible model endpoints, including free development endpoints for selected models.
 
 A curated model list marks what we've verified for tool-calling work. Adding any model string works at your own risk.
 

coworker/providers/capabilities.py

@@ -29,6 +29,15 @@ def capabilities_for(model: str) -> ModelCapabilities:
             tools=True, vision=False, parallel_tool_calls=False, streaming=True
         )
 
+    # NVIDIA NIM (custom-added ids; curated ones answered from the matrix above). The catalog
+    # spans many labs' checkpoints (Nemotron, Llama, DeepSeek, Qwen, …) with mixed tool-calling
+    # reliability across them, so stay conservative on parallel calls until a specific model is
+    # probed and added to the matrix — mirrors the Bedrock/Vertex custom-id posture below.
+    if provider == "nvidia":
+        return ModelCapabilities(
+            tools=True, vision=False, parallel_tool_calls=False, streaming=True
+        )
+
     # Cloud-account providers (custom-added ids; curated ones answered from the matrix).
     # The family segment decides: Claude keeps its native capabilities; everything else
     # stays conservative until probed (Converse tool calling works across families, but

coworker/providers/matrix.py

@@ -16,8 +16,8 @@
 showing a made-up denominator. Values entered 2026-07-28 from vendor docs; verify alongside
 the id refresh.
 
-Resellers: Together + Fireworks + OpenRouter. TODO: add Groq entries here AND its
-descriptor in ``registry.py`` once the current provider surface is tested — deliberately
+Resellers: Together + Fireworks + OpenRouter + NVIDIA NIM. TODO: add Groq entries here AND
+its descriptor in ``registry.py`` once the current provider surface is tested — deliberately
 deferred to bound how much needs verifying at once.
 """
 
@@ -110,6 +110,16 @@ class ModelEntry:
         "Mistral Large · Mistral", _AGENTIC, 128_000
     ),
     # -- resellers (their model namespaces, verbatim) -----------------------------
+    "nvidia:nvidia/nemotron-3-nano-30b-a3b": ModelEntry(
+        "Nemotron 3 Nano 30B · NVIDIA NIM",
+        ModelCapabilities(
+            tools=True,
+            vision=False,
+            parallel_tool_calls=False,
+            streaming=True,
+        ),
+        262_144,
+    ),
     "together:thinkingmachines/Inkling": ModelEntry("Inkling · via Together"),
     "together:zai-org/GLM-5.2": ModelEntry("GLM-5.2 · via Together", _AGENTIC, 128_000),
     # Kimi K3 on Together (landed late July 2026): 1M window, native vision; PDFs
@@ -157,6 +167,18 @@ class ModelEntry:
     "openrouter:meta-llama/llama-4-maverick": ModelEntry(
         "Llama 4 Maverick · via OpenRouter", _AGENTIC, 1_000_000
     ),
+    # NVIDIA NIM ids use the catalog's own `<publisher>/<model>` namespace (checked against
+    # the NIM catalog 2026-07-04). Nemotron Super 3 120B is NVIDIA's own flagship — the reason
+    # to reach for NIM over another reseller when it's on the free tier.
+    "nvidia:nvidia/nemotron-super-3-120b": ModelEntry(
+        "Nemotron Super 3 120B · via NVIDIA NIM", _AGENTIC, 128_000
+    ),
+    "nvidia:deepseek-ai/deepseek-v4-pro": ModelEntry(
+        "DeepSeek V4 Pro · via NVIDIA NIM", _AGENTIC, 128_000
+    ),
+    "nvidia:meta/llama-4-maverick-17b-128e-instruct": ModelEntry(
+        "Llama 4 Maverick · via NVIDIA NIM", _AGENTIC, 1_000_000
+    ),
     # -- cloud accounts (models running in the user's own AWS/GCP) ----------------
     # Bedrock ids carry a family segment (claude/ → native Anthropic path, other/ →
     # Converse) plus AWS's own `-v<n>:<m>` version suffix. Some regions require the

coworker/providers/registry.py

@@ -550,6 +550,21 @@ def _compat(
         recommended_model="z-ai/glm-5.2",
         env_key="OPENROUTER_API_KEY",
     ),
+    # NVIDIA NIM: NVIDIA's own hosted catalog of 100+ open-weight models (its own Nemotron
+    # line alongside Llama, DeepSeek, Qwen, Mistral, and others), all behind one key and one
+    # OpenAI-compatible endpoint. Free API keys (no credit card) are issued at build.nvidia.com,
+    # which is what makes this worth listing next to the paid resellers above.
+    _compat(
+        "nvidia",
+        "NVIDIA NIM",
+        base_url="https://integrate.api.nvidia.com/v1",
+        recommended_model="nvidia/nemotron-3-nano-30b-a3b",
+        env_key="NVIDIA_API_KEY",
+        endpoint_help=(
+            "Prefilled with NVIDIA's hosted NIM endpoint. "
+            "You can replace it with a self-hosted NIM or compatible proxy."
+        ),
+    ),
     ProviderDescriptor(
         name="ollama",
         title="Ollama (local models)",