Agent Zero AI framework
model api_base, litellm finalizing
frdeldfc7de05145c
diff --git a/agent.py b/agent.pyindex b92eed2d0..1c83e3c46 100644--- a/agent.py+++ b/agent.py@@ -206,6 +206,7 @@ class AgentContext:class ModelConfig:provider: models.ModelProvidername: str+ api_base: str = ""ctx_length: int = 0limit_requests: int = 0limit_input: int = 0@@ -581,23 +582,29 @@ class Agent:return models.get_chat_model(self.config.chat_model.provider,self.config.chat_model.name,- **self.config.chat_model.kwargs,+ **self._get_model_kwargs(self.config.chat_model),)def get_utility_model(self):return models.get_chat_model(self.config.utility_model.provider,self.config.utility_model.name,- **self.config.utility_model.kwargs,+ **self._get_model_kwargs(self.config.utility_model),)def get_embedding_model(self):return models.get_embedding_model(self.config.embeddings_model.provider,self.config.embeddings_model.name,- **self.config.embeddings_model.kwargs,+ **self._get_model_kwargs(self.config.embeddings_model),)+ def _get_model_kwargs(self, model_config: ModelConfig):+ kwargs = model_config.kwargs.copy() or {}+ if model_config.api_base and "api_base" not in kwargs:+ kwargs["api_base"] = model_config.api_base+ return kwargs+async def call_utility_model(self,system: str,diff --git a/initialize.py b/initialize.pyindex 3c7b3f849..8fbd94b31 100644--- a/initialize.py+++ b/initialize.py@@ -29,6 +29,7 @@ def initialize_agent():chat_llm = ModelConfig(provider=models.ModelProvider[current_settings["chat_model_provider"]],name=current_settings["chat_model_name"],+ api_base=current_settings["chat_model_api_base"],ctx_length=current_settings["chat_model_ctx_length"],vision=current_settings["chat_model_vision"],limit_requests=current_settings["chat_model_rl_requests"],@@ -41,6 +42,7 @@ def initialize_agent():utility_llm = ModelConfig(provider=models.ModelProvider[current_settings["util_model_provider"]],name=current_settings["util_model_name"],+ api_base=current_settings["util_model_api_base"],ctx_length=current_settings["util_model_ctx_length"],limit_requests=current_settings["util_model_rl_requests"],limit_input=current_settings["util_model_rl_input"],@@ -51,6 +53,7 @@ def initialize_agent():embedding_llm = ModelConfig(provider=models.ModelProvider[current_settings["embed_model_provider"]],name=current_settings["embed_model_name"],+ api_base=current_settings["embed_model_api_base"],limit_requests=current_settings["embed_model_rl_requests"],kwargs=_normalize_model_kwargs(current_settings["embed_model_kwargs"]),)@@ -58,6 +61,7 @@ def initialize_agent():browser_llm = ModelConfig(provider=models.ModelProvider[current_settings["browser_model_provider"]],name=current_settings["browser_model_name"],+ api_base=current_settings["browser_model_api_base"],vision=current_settings["browser_model_vision"],kwargs=_normalize_model_kwargs(current_settings["browser_model_kwargs"]),)diff --git a/models.py b/models.pyindex 41a70934b..d487c940c 100644--- a/models.py+++ b/models.py@@ -59,19 +59,18 @@ class ModelType(Enum):class ModelProvider(Enum):ANTHROPIC = "Anthropic"- CHUTES = "Chutes"DEEPSEEK = "DeepSeek"- GOOGLE = "Google"+ GEMINI = "Google"GROQ = "Groq"HUGGINGFACE = "HuggingFace"- LMSTUDIO = "LM Studio"- MISTRALAI = "Mistral AI"+ LM_STUDIO = "LM Studio"+ MISTRAL = "Mistral AI"OLLAMA = "Ollama"OPENAI = "OpenAI"AZURE = "OpenAI Azure"OPENROUTER = "OpenRouter"SAMBANOVA = "Sambanova"- OTHER = "Other"+ OTHER = "Other OpenAI compatible"class ChatChunk(TypedDict):@@ -84,42 +83,6 @@ class ChatChunk(TypedDict):rate_limiters: dict[str, RateLimiter] = {}-def configure_litellm_environment():- env_mappings = {- "API_KEY_OPENAI": "OPENAI_API_KEY",- "API_KEY_ANTHROPIC": "ANTHROPIC_API_KEY",- "API_KEY_GROQ": "GROQ_API_KEY",- "API_KEY_GOOGLE": "GOOGLE_API_KEY",- "API_KEY_MISTRAL": "MISTRAL_API_KEY",- "API_KEY_OLLAMA": "OLLAMA_API_KEY",- "API_KEY_HUGGINGFACE": "HUGGINGFACE_API_KEY",- "API_KEY_OPENAI_AZURE": "AZURE_AI_API_KEY",- "API_KEY_DEEPSEEK": "DEEPSEEK_API_KEY",- "API_KEY_SAMBANOVA": "SAMBANOVA_API_KEY",- "API_KEY_GOOGLE": "GEMINI_API_KEY",- }- base_url_mappings = {- "OPENAI_BASE_URL": "OPENAI_API_BASE",- "ANTHROPIC_BASE_URL": "ANTHROPIC_API_BASE",- "GROQ_BASE_URL": "GROQ_API_BASE",- "GOOGLE_BASE_URL": "GOOGLE_API_BASE",- "MISTRAL_BASE_URL": "MISTRAL_API_BASE",- "OLLAMA_BASE_URL": "OLLAMA_API_BASE",- "HUGGINGFACE_BASE_URL": "HUGGINGFACE_API_BASE",- "AZURE_BASE_URL": "AZURE_AI_API_BASE",- "DEEPSEEK_BASE_URL": "DEEPSEEK_API_BASE",- "SAMBANOVA_BASE_URL": "SAMBANOVA_API_BASE",- }- for a0, llm in env_mappings.items():- val = dotenv.get_dotenv_value(a0)- if val and not os.getenv(llm):- os.environ[llm] = val- for a0_base, llm_base in base_url_mappings.items():- val = dotenv.get_dotenv_value(a0_base)- if val and not os.getenv(llm_base):- os.environ[llm_base] = val--def get_api_key(service: str) -> str:return (dotenv.get_dotenv_value(f"API_KEY_{service.upper()}")@@ -140,26 +103,6 @@ def get_rate_limiter(return limiter-def _parse_chunk(chunk: Any) -> ChatChunk:- delta = chunk["choices"][0].get("delta", {})- message = chunk["choices"][0].get("model_extra", {}).get("message", {})- response_delta = (- delta.get("content", "")- if isinstance(delta, dict)- else getattr(delta, "content", "")- ) or (- message.get("content", "")- if isinstance(message, dict)- else getattr(message, "content", "")- )- reasoning_delta = (- delta.get("reasoning_content", "")- if isinstance(delta, dict)- else getattr(delta, "reasoning_content", "")- )- return ChatChunk(reasoning_delta=reasoning_delta, response_delta=response_delta)--class LiteLLMChatWrapper(SimpleChatModel):model_name: strprovider: str@@ -284,7 +227,7 @@ class LiteLLMChatWrapper(SimpleChatModel):self,system_message="",user_message="",- messages: List[BaseMessage]|None = None,+ messages: List[BaseMessage] | None = None,response_callback: Callable[[str, str], Awaitable[None]] | None = None,reasoning_callback: Callable[[str, str], Awaitable[None]] | None = None,tokens_callback: Callable[[str, int], Awaitable[None]] | None = None,@@ -424,66 +367,86 @@ def _get_litellm_chat(provider_name: str = "",**kwargs: Any,):- provider_name = provider_name.lower()-- configure_litellm_environment()- # Use original provider name for API key lookup, fallback to mapped provider name+ # use api key from kwargs or envapi_key = kwargs.pop("api_key", None) or get_api_key(provider_name)- # litellm will pick up base_url from env. We just need to control the api_key.- # base_url = dotenv.get_dotenv_value(f"{provider_name.upper()}_BASE_URL")-- # If a base_url is set, ensure api_key is not passed to litellm- # > remove, this can be handled by api_key=None- # if base_url:- # if "api_key" in kwargs:- # del kwargs["api_key"]- # Only pass API key if no base_url is set and key is not a placeholder+ # Only pass API key if key is not a placeholderif api_key and api_key not in ("None", "NA"):kwargs["api_key"] = api_key- # for openrouter add app reference- if provider_name == "openrouter":- kwargs["extra_headers"] = {- "HTTP-Referer": "https://agent-zero.ai",- "X-Title": "Agent Zero",- }-- return cls(model=model_name, provider=provider_name, **kwargs)+ provider_name, model_name, kwargs = _adjust_call_args(+ provider_name, model_name, kwargs+ )+ return cls(provider=provider_name, model=model_name, **kwargs)-def get_litellm_embedding(model_name: str, provider: str, **kwargs: Any):+def _get_litellm_embedding(model_name: str, provider_name: str, **kwargs: Any):# Check if this is a local sentence-transformers model- if provider == "huggingface" and model_name.startswith("sentence-transformers/"):+ if provider_name == "huggingface" and model_name.startswith(+ "sentence-transformers/"+ ):# Use local sentence-transformers instead of LiteLLM for local models- return LocalSentenceTransformerWrapper(provider=provider, model=model_name, **kwargs)-- configure_litellm_environment()- # Use original provider name for API key lookup, fallback to mapped provider name- api_key = kwargs.pop("api_key", None) or get_api_key(provider)+ provider_name, model_name, kwargs = _adjust_call_args(+ provider_name, model_name, kwargs+ )+ return LocalSentenceTransformerWrapper(+ provider=provider_name, model=model_name, **kwargs+ )- # litellm will pick up base_url from env. We just need to control the api_key.- # base_url = dotenv.get_dotenv_value(f"{provider.upper()}_BASE_URL")+ # use api key from kwargs or env+ api_key = kwargs.pop("api_key", None) or get_api_key(provider_name)- # If a base_url is set, ensure api_key is not passed to litellm- # > remove, this can be handled by api_key=None- # if base_url:- # if "api_key" in kwargs:- # del kwargs["api_key"]- # Only pass API key if no base_url is set and key is not a placeholder+ # Only pass API key if key is not a placeholderif api_key and api_key not in ("None", "NA"):kwargs["api_key"] = api_key- return LiteLLMEmbeddingWrapper(model=model_name, provider=provider, **kwargs)+ provider_name, model_name, kwargs = _adjust_call_args(+ provider_name, model_name, kwargs+ )+ return LiteLLMEmbeddingWrapper(model=model_name, provider=provider_name, **kwargs)+++def _parse_chunk(chunk: Any) -> ChatChunk:+ delta = chunk["choices"][0].get("delta", {})+ message = chunk["choices"][0].get("model_extra", {}).get("message", {})+ response_delta = (+ delta.get("content", "")+ if isinstance(delta, dict)+ else getattr(delta, "content", "")+ ) or (+ message.get("content", "")+ if isinstance(message, dict)+ else getattr(message, "content", "")+ )+ reasoning_delta = (+ delta.get("reasoning_content", "")+ if isinstance(delta, dict)+ else getattr(delta, "reasoning_content", "")+ )+ return ChatChunk(reasoning_delta=reasoning_delta, response_delta=response_delta)+++def _adjust_call_args(provider_name: str, model_name: str, kwargs: dict):+ # for openrouter add app reference+ if provider_name == "openrouter":+ kwargs["extra_headers"] = {+ "HTTP-Referer": "https://agent-zero.ai",+ "X-Title": "Agent Zero",+ }++ # remap other to openai for litellm+ if provider_name == "other":+ provider_name = "openai"++ return provider_name, model_name, kwargsdef get_model(type: ModelType, provider: ModelProvider, name: str, **kwargs: Any):provider_name = provider.name.lower()- kwargs = _normalize_chat_kwargs(provider, kwargs)if type == ModelType.CHAT:return _get_litellm_chat(LiteLLMChatWrapper, name, provider_name, **kwargs)elif type == ModelType.EMBEDDING:- return get_litellm_embedding(name, provider_name, **kwargs)+ return _get_litellm_embedding(name, provider_name, **kwargs)else:raise ValueError(f"Unsupported model type: {type}")@@ -491,8 +454,7 @@ def get_model(type: ModelType, provider: ModelProvider, name: str, **kwargs: Anydef get_chat_model(provider: ModelProvider, name: str, **kwargs: Any) -> LiteLLMChatWrapper:- provider_name = _get_litellm_provider(provider)- kwargs = _normalize_chat_kwargs(provider, kwargs)+ provider_name = provider.name.lower()model = _get_litellm_chat(LiteLLMChatWrapper, name, provider_name, **kwargs)return model@@ -501,7 +463,6 @@ def get_browser_model(provider: ModelProvider, name: str, **kwargs: Any) -> BrowserCompatibleChatWrapper:provider_name = provider.name.lower()- kwargs = _normalize_chat_kwargs(provider, kwargs)model = _get_litellm_chat(BrowserCompatibleChatWrapper, name, provider_name, **kwargs)@@ -512,30 +473,5 @@ def get_embedding_model(provider: ModelProvider, name: str, **kwargs: Any) -> LiteLLMEmbeddingWrapper | LocalSentenceTransformerWrapper:provider_name = provider.name.lower()- kwargs = _normalize_embedding_kwargs(kwargs)- model = get_litellm_embedding(name, provider_name, **kwargs)+ model = _get_litellm_embedding(name, provider_name, **kwargs)return model---def _normalize_chat_kwargs(provider: ModelProvider, kwargs: Any) -> Any:- # this prevents using openai api key for other providers- if provider == ModelProvider.OTHER:- if "api_key" not in kwargs:- kwargs["api_key"] = "None"- return kwargs---def _normalize_embedding_kwargs(kwargs: Any) -> Any:- return kwargs---def _get_litellm_provider(provider: ModelProvider) -> str:- name = provider.name.lower()-- # exceptions- if name == "google":- name = "gemini"- elif name == "other":- name = "openai"-- return namediff --git a/python/helpers/settings.py b/python/helpers/settings.pyindex 5aa0a64ca..1db6d4e94 100644--- a/python/helpers/settings.py+++ b/python/helpers/settings.py@@ -17,6 +17,7 @@ class Settings(TypedDict):chat_model_provider: strchat_model_name: str+ chat_model_api_base: strchat_model_kwargs: dict[str, str]chat_model_ctx_length: intchat_model_ctx_history: float@@ -27,6 +28,7 @@ class Settings(TypedDict):util_model_provider: strutil_model_name: str+ util_model_api_base: strutil_model_kwargs: dict[str, str]util_model_ctx_length: intutil_model_ctx_input: float@@ -36,12 +38,14 @@ class Settings(TypedDict):embed_model_provider: strembed_model_name: str+ embed_model_api_base: strembed_model_kwargs: dict[str, str]embed_model_rl_requests: intembed_model_rl_input: intbrowser_model_provider: strbrowser_model_name: str+ browser_model_api_base: strbrowser_model_vision: boolbrowser_model_kwargs: dict[str, str]@@ -141,6 +145,16 @@ def convert_out(settings: Settings) -> SettingsOutput:})+ chat_model_fields.append(+ {+ "id": "chat_model_api_base",+ "title": "Chat model API base URL",+ "description": "API base URL for main chat model. Leave empty for default. Only relevant for Azure, local and custom (other) providers.",+ "type": "text",+ "value": settings["chat_model_api_base"],+ }+ )+chat_model_fields.append({"id": "chat_model_ctx_length",@@ -208,8 +222,7 @@ def convert_out(settings: Settings) -> SettingsOutput:{"id": "chat_model_kwargs","title": "Chat model additional parameters",- "description": """Any other parameters supported by the model. Format is KEY=VALUE on individual lines, just like .env file.- For OpenAI compatible providers not listed here, select 'other' and specify api_base=https://... and api_key=... as additional parameters.""",+ "description": "Any other parameters supported by <a href='https://docs.litellm.ai/docs/set_keys' target='_blank'>LiteLLM</a>. Format is KEY=VALUE on individual lines, just like .env file.","type": "textarea","value": _dict_to_env(settings["chat_model_kwargs"]),}@@ -245,6 +258,16 @@ def convert_out(settings: Settings) -> SettingsOutput:})+ util_model_fields.append(+ {+ "id": "util_model_api_base",+ "title": "Utility model API base URL",+ "description": "API base URL for utility model. Leave empty for default. Only relevant for Azure, local and custom (other) providers.",+ "type": "text",+ "value": settings["util_model_api_base"],+ }+ )+util_model_fields.append({"id": "util_model_rl_requests",@@ -279,8 +302,7 @@ def convert_out(settings: Settings) -> SettingsOutput:{"id": "util_model_kwargs","title": "Utility model additional parameters",- "description": """Any other parameters supported by the model. Format is KEY=VALUE on individual lines, just like .env file.- For OpenAI compatible providers not listed here, select 'other' and specify api_base=https://... and api_key=... as additional parameters.""",+ "description": "Any other parameters supported by <a href='https://docs.litellm.ai/docs/set_keys' target='_blank'>LiteLLM</a>. Format is KEY=VALUE on individual lines, just like .env file.","type": "textarea","value": _dict_to_env(settings["util_model_kwargs"]),}@@ -316,6 +338,16 @@ def convert_out(settings: Settings) -> SettingsOutput:})+ embed_model_fields.append(+ {+ "id": "embed_model_api_base",+ "title": "Embedding model API base URL",+ "description": "API base URL for embedding model. Leave empty for default. Only relevant for Azure, local and custom (other) providers.",+ "type": "text",+ "value": settings["embed_model_api_base"],+ }+ )+embed_model_fields.append({"id": "embed_model_rl_requests",@@ -340,8 +372,7 @@ def convert_out(settings: Settings) -> SettingsOutput:{"id": "embed_model_kwargs","title": "Embedding model additional parameters",- "description": """Any other parameters supported by the model. Format is KEY=VALUE on individual lines, just like .env file.- For OpenAI compatible providers not listed here, select 'other' and specify api_base=https://... and api_key=... as additional parameters.""",+ "description": "Any other parameters supported by <a href='https://docs.litellm.ai/docs/set_keys' target='_blank'>LiteLLM</a>. Format is KEY=VALUE on individual lines, just like .env file.","type": "textarea","value": _dict_to_env(settings["embed_model_kwargs"]),}@@ -391,7 +422,7 @@ def convert_out(settings: Settings) -> SettingsOutput:{"id": "browser_model_kwargs","title": "Web Browser model additional parameters",- "description": "Any other parameters supported by the model. Format is KEY=VALUE on individual lines, just like .env file.",+ "description": "Any other parameters supported by <a href='https://docs.litellm.ai/docs/set_keys' target='_blank'>LiteLLM</a>. Format is KEY=VALUE on individual lines, just like .env file.","type": "textarea","value": _dict_to_env(settings["browser_model_kwargs"]),}@@ -472,26 +503,9 @@ def convert_out(settings: Settings) -> SettingsOutput:# api keys model sectionapi_keys_fields: list[SettingsField] = []- api_keys_fields.append(_get_api_key_field(settings, "openai", "OpenAI API Key"))- api_keys_fields.append(- _get_api_key_field(settings, "anthropic", "Anthropic API Key")- )- api_keys_fields.append(_get_api_key_field(settings, "chutes", "Chutes API Key"))- api_keys_fields.append(_get_api_key_field(settings, "deepseek", "DeepSeek API Key"))- api_keys_fields.append(_get_api_key_field(settings, "google", "Google API Key"))- api_keys_fields.append(_get_api_key_field(settings, "groq", "Groq API Key"))- api_keys_fields.append(- _get_api_key_field(settings, "huggingface", "HuggingFace API Key")- )- api_keys_fields.append(- _get_api_key_field(settings, "mistralai", "MistralAI API Key")- )- api_keys_fields.append(- _get_api_key_field(settings, "openrouter", "OpenRouter API Key")- )- api_keys_fields.append(- _get_api_key_field(settings, "sambanova", "Sambanova API Key")- )++ for provider in ModelProvider:+ api_keys_fields.append(_get_api_key_field(settings, provider.name.lower(), provider.value))api_keys_section: SettingsSection = {"id": "api_keys",@@ -965,6 +979,7 @@ def get_default_settings() -> Settings:version=_get_version(),chat_model_provider=ModelProvider.OPENROUTER.name,chat_model_name="openai/gpt-4.1",+ chat_model_api_base="",chat_model_kwargs={"temperature": "0"},chat_model_ctx_length=100000,chat_model_ctx_history=0.7,@@ -974,6 +989,7 @@ def get_default_settings() -> Settings:chat_model_rl_output=0,util_model_provider=ModelProvider.OPENROUTER.name,util_model_name="openai/gpt-4.1-nano",+ util_model_api_base="",util_model_ctx_length=100000,util_model_ctx_input=0.7,util_model_kwargs={"temperature": "0"},@@ -982,11 +998,13 @@ def get_default_settings() -> Settings:util_model_rl_output=0,embed_model_provider=ModelProvider.HUGGINGFACE.name,embed_model_name="sentence-transformers/all-MiniLM-L6-v2",+ embed_model_api_base="",embed_model_kwargs={},embed_model_rl_requests=0,embed_model_rl_input=0,browser_model_provider=ModelProvider.OPENROUTER.name,browser_model_name="openai/gpt-4.1",+ browser_model_api_base="",browser_model_vision=True,browser_model_kwargs={"temperature": "0"},api_keys={},diff --git a/python/tools/browser_agent.py b/python/tools/browser_agent.pyindex cf6d300f2..79be1cedd 100644--- a/python/tools/browser_agent.py+++ b/python/tools/browser_agent.py@@ -57,7 +57,13 @@ class State:viewport={"width": 1024, "height": 2048},args=["--headless=new"],# Use a unique user data directory to avoid conflicts- user_data_dir=str(Path.home() / ".config" / "browseruse" / "profiles" / f"agent_{self.agent.context.id}"),+ user_data_dir=str(+ Path.home()+ / ".config"+ / "browseruse"+ / "profiles"+ / f"agent_{self.agent.context.id}"+ ),))@@ -119,11 +125,10 @@ class State:)return result-model = models.get_browser_model(provider=self.agent.config.browser_model.provider,name=self.agent.config.browser_model.name,- **self.agent.config.browser_model.kwargs,+ **self.agent._get_model_kwargs(self.agent.config.browser_model),)try:@@ -140,7 +145,9 @@ class State:# available_file_paths=[],)except Exception as e:- raise Exception(f"Browser agent initialization failed. This might be due to model compatibility issues. Error: {e}") from e+ raise Exception(+ f"Browser agent initialization failed. This might be due to model compatibility issues. Error: {e}"+ ) from eself.iter_no = get_iter_no(self.agent)@@ -298,13 +305,17 @@ class BrowserAgent(Tool):f"Task reached step limit without completion. Last page: {current_url}. "f"The browser agent may need clearer instructions on when to finish.")-+# update the log (without screenshot path here, user can click)self.log.update(answer=answer_text)# add screenshot to the answer if we have it- if self.log.kvps and "screenshot" in self.log.kvps and self.log.kvps['screenshot']:- path = self.log.kvps['screenshot'].split('//', 1)[-1].split('&', 1)[0]+ if (+ self.log.kvps+ and "screenshot" in self.log.kvps+ and self.log.kvps["screenshot"]+ ):+ path = self.log.kvps["screenshot"].split("//", 1)[-1].split("&", 1)[0]answer_text += f"\n\nScreenshot: {path}"# respond (with screenshot path)@@ -416,7 +427,9 @@ def get_use_agent_log(use_agent: browser_use.Agent | None):if item.success:short_log.append(f"✅ Done")else:- short_log.append(f"❌ Error: {item.error or item.extracted_content or 'Unknown error'}")+ short_log.append(+ f"❌ Error: {item.error or item.extracted_content or 'Unknown error'}"+ )# progress messageselse: