feat: Add Forge as LLM provider

agentflow/.env.template

@@ -23,6 +23,8 @@ GOOGLE_API_KEY=<your-api-key-here>
 ANTHROPIC_API_KEY=<your-api-key-here>      # Anthropic LLM
 DEEPSEEK_API_KEY=<your-api-key-here>       # DeepSeek LLM
 XAI_API_KEY=<your-api-key-here>            # Grok (xAI) LLM
+FORGE_API_KEY=<your-api-key-here>          # Forge LLM router
+FORGE_API_BASE=https://api.forge.tensorblock.co/v1  # Optional Forge base URL override
 
 # === Azure OpenAI Configuration ===
 AZURE_OPENAI_API_KEY=<your-api-key-here>   # Azure OpenAI API Key

agentflow/agentflow/engine/factory.py

@@ -19,6 +19,29 @@ def create_llm_engine(model_string: str, use_cache: bool = False, is_multimodal:
 
     print(f"creating llm engine {model_string} with: is_multimodal: {is_multimodal}, kwargs: {kwargs}")
 
+    model_string_lower = model_string.lower()
+
+    # === Forge (OpenAI-compatible router) ===
+    if model_string_lower.startswith("forge-"):
+        from .openai import ChatOpenAI
+
+        if "FORGE_API_KEY" not in os.environ:
+            raise ValueError("Please set the FORGE_API_KEY environment variable.")
+
+        model_string = model_string[len("forge-"):]
+        config = {
+            "model_string": model_string,
+            "use_cache": use_cache,
+            "is_multimodal": is_multimodal,
+            "temperature": kwargs.get("temperature", 0.7),
+            "top_p": kwargs.get("top_p", 0.9),
+            "frequency_penalty": kwargs.get("frequency_penalty", 0.5),
+            "presence_penalty": kwargs.get("presence_penalty", 0.5),
+            "api_key": os.getenv("FORGE_API_KEY"),
+            "base_url": os.getenv("FORGE_API_BASE") or "https://api.forge.tensorblock.co/v1",
+        }
+        return ChatOpenAI(**config)
+
     # === Azure OpenAI ===
     if "azure" in model_string:
         from .azure import ChatAzureOpenAI
@@ -196,4 +219,4 @@ def create_llm_engine(model_string: str, use_cache: bool = False, is_multimodal:
             "For Ollama models, use 'ollama-'. "
             "For other custom engines, you can edit the factory.py file and add its interface file. "
             "Your pull request will be warmly welcomed!"
-        )
+        )

agentflow/agentflow/engine/openai.py

@@ -63,6 +63,8 @@ def __init__(
         system_prompt=DEFAULT_SYSTEM_PROMPT,
         is_multimodal: bool=False,
         use_cache: bool=True, # disable cache for now
+        api_key: str | None = None,
+        base_url: str | None = None,
         **kwargs):
         """
         :param model_string:
@@ -87,12 +89,15 @@ def __init__(
             os.makedirs(self.image_cache_dir, exist_ok=True)
             super().__init__(cache_path=cache_path)
 
-        if os.getenv("OPENAI_API_KEY") is None:
+        resolved_api_key = api_key or os.getenv("OPENAI_API_KEY")
+        if not resolved_api_key:
             raise ValueError("Please set the OPENAI_API_KEY environment variable if you'd like to use OpenAI models.")
-
-        self.client = OpenAI(
-            api_key=os.getenv("OPENAI_API_KEY"),
-        )
+
+        client_kwargs = {"api_key": resolved_api_key}
+        if base_url:
+            client_kwargs["base_url"] = base_url
+
+        self.client = OpenAI(**client_kwargs)
 
 
     @retry(wait=wait_random_exponential(min=1, max=5), stop=stop_after_attempt(5))

assets/doc/api_key.md

@@ -73,6 +73,16 @@ This guide provides detailed instructions on how to obtain API keys for all LLM
 
 > **Important Note for Qwen Models**: Together AI offers both Turbo (quantized) and standard (non-quantized) versions of Qwen models. For best performance and accuracy, we recommend using the **non-quantized versions** (e.g., `Qwen/Qwen2.5-7B-Instruct` instead of `Qwen/Qwen2.5-7B-Instruct-Turbo`). The Turbo versions are faster but may have reduced quality due to quantization.
 
+---
+
+## 5. Forge API Key
+
+**Purpose**: Access Forge's OpenAI-compatible router with model strings like `OpenAI/gpt-4o-mini`.
+
+**Environment variables**:
+- `FORGE_API_KEY`
+- `FORGE_API_BASE` (optional, defaults to `https://api.forge.tensorblock.co/v1`)
+
 
 ## Important Notes
 
@@ -94,3 +104,4 @@ This guide provides detailed instructions on how to obtain API keys for all LLM
 | GOOGLE_API_KEY | Gemini models | [aistudio.google.com](https://aistudio.google.com/) | [Docs](https://ai.google.dev/gemini-api/docs/models/gemini) |
 | DASHSCOPE_API_KEY | Qwen models | [dashscope.console.aliyun.com](https://dashscope.console.aliyun.com/) | [Docs](https://help.aliyun.com/zh/dashscope/developer-reference/model-square) |
 | TOGETHER_API_KEY | Qwen & open-source models | [together.ai](https://www.together.ai/) | [Docs](https://docs.together.ai/docs/inference-models) |
+| FORGE_API_KEY | Forge router | [forge.tensorblock.co](https://forge.tensorblock.co) | [Docs](https://forge.tensorblock.co) |

assets/doc/llm_engine.md

@@ -7,6 +7,7 @@ We support a broad range of LLM engines for agents and tools in [`factory.py`](.
 | vLLM | `vllm-Qwen/Qwen2.5-7B-Instruct` | Various vLLM-supported models (e.g., `Qwen2.5-7B-Instruct`, `Qwen2.5-VL-3B-Instruct`). Supports local checkpoint models for customization and local inference. | [vLLM Models](https://docs.vllm.ai/en/latest/models/supported_models.html) |
 | DashScope (Qwen) | `dashscope-qwen2.5-7b-instruct` | Qwen models via Alibaba Cloud DashScope API | [DashScope Models](https://help.aliyun.com/zh/model-studio/getting-started/models) |
 | OpenAI | `gpt-4o`, `o1-mini` | `gpt-4-turbo`, `gpt-4o`, `gpt-4o-mini`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`, `gpt-3.5-turbo`, `gpt-4`, `o1`, `o1-mini`, `o3`, `o3-mini`, `o1-pro`, `o4-mini` | [OpenAI Models](https://platform.openai.com/docs/models) |
+| Forge | `forge-OpenAI/gpt-4o-mini` | Any OpenAI-compatible model via Forge router (`Provider/model-name`) | [Forge](https://forge.tensorblock.co) |
 | Azure OpenAI | `azure-gpt-4o` | `gpt-4o`, `gpt-4o-mini`, `gpt-4.1`, `gpt-4.1-mini`, `gpt-4.1-nano`, `gpt-3.5-turbo`, `gpt-4`, `o1`, `o1-mini`, `o3`, `o3-mini`, `o1-pro`, `o4-mini` | [Azure OpenAI Models](https://learn.microsoft.com/en-us/azure/ai-services/openai/reference#models) |
 | Anthropic | `claude-3-5-sonnet-20241022` | `claude-3-haiku-20240307`, `claude-3-sonnet-20240229`, `claude-3-opus-20240229`, `claude-3-5-sonnet-20240620`, `claude-3-5-sonnet-20241022`, `claude-3-5-haiku-20241022`, `claude-3-7-sonnet-20250219` | [Anthropic Models](https://docs.anthropic.com/en/docs/about-claude/models/all-models) |
 | TogetherAI | `together-meta-llama/Llama-3-70b-chat-hf` | Most models including `meta-llama/Llama-4-Scout-17B-16E-Instruct`, `Qwen/QwQ-32B`, `Qwen/Qwen2-VL-72B-Instruct`, `meta-llama/Llama-3-70b-chat-hf`, `Qwen/Qwen2-72B-Instruct` | [TogetherAI Models](https://api.together.ai/models) |