Requesty provider

.env.example

@@ -15,6 +15,7 @@ GROQ_API_KEY=
 OPENAI_API_KEY=
 OPENROUTER_API_KEY=
 PERPLEXITY_API_KEY=
+REQUESTY_API_KEY=
 TOGETHER_API_KEY=
 
 # --- AWS Bedrock ---

AutoCua/android/controller/tool/web/service.rs

@@ -36,7 +36,7 @@ use crate::agent::android_use::service::read_text;
 /// Providers with no native web-search API: the query goes to the browser
 /// agent on the same provider+model instead (see web_agent.rs). Explicit
 /// membership, not a catch-all: a typo'd provider must still fail fast below.
-const BROWSER_FALLBACK_PROVIDERS: [&str; 3] = ["together", "cerebras", "aws"];
+const BROWSER_FALLBACK_PROVIDERS: [&str; 4] = ["together", "cerebras", "requesty", "aws"];
 
 /// AutoCua/android/controller/tool/web: this tool's folder.
 fn web_dir(py: Python<'_>) -> PyResult<PathBuf> {

AutoCua/android/controller/tool/web/web_agent.rs

@@ -1,6 +1,6 @@
 //! web_agent.py: the browser-agent fallback for the `web` tool. Providers
-//! with no native web-search API (together, cerebras, aws) still have to
-//! honour the `web` action, so the query goes to the browser agent
+//! with no native web-search API (together, cerebras, requesty, aws) still
+//! have to honour the `web` action, so the query goes to the browser agent
 //! (AutoCua/web, the Rust/CDP Chrome driver) running on the SAME provider
 //! and model; its `done` summary is the web result. The task text below is
 //! what turns that summary into a proper report rather than a one-line
@@ -35,9 +35,10 @@ const SHORT_REPORT_CHARS: usize = 200;
 /// The runtime (frontend) key reaches the browser agent through its env: the
 /// web agent's __main__ has no --api_key flag, and its .env loader never
 /// overrides a var that is already set.
-const PROVIDER_ENV_KEY: [(&str, &str); 3] = [
+const PROVIDER_ENV_KEY: [(&str, &str); 4] = [
     ("together", "TOGETHER_API_KEY"),
     ("cerebras", "CEREBRAS_API_KEY"),
+    ("requesty", "REQUESTY_API_KEY"),
     ("aws", "AWS_API_KEY"),
 ];
 

AutoCua/ios/controller/tool/web/service.py

@@ -17,7 +17,7 @@ def _web_search(provider: str):
 # Providers with no native web-search API: the query goes to the browser
 # agent on the same provider+model instead (see web_agent.py). Explicit
 # membership, not a catch-all - a typo'd provider must still fail fast below.
-_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "aws"})
+_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "requesty", "aws"})
 
 
 class WebService:

AutoCua/ios/controller/tool/web/web_agent.py

@@ -48,7 +48,7 @@
 # __main__ has no --api_key flag, and its .env loader never overrides a var
 # that is already set.
 _PROVIDER_ENV_KEY = {"together": "TOGETHER_API_KEY", "cerebras": "CEREBRAS_API_KEY",
-                     "aws": "AWS_API_KEY"}
+                     "requesty": "REQUESTY_API_KEY", "aws": "AWS_API_KEY"}
 # A `done` value shorter than this is almost certainly an activity log, not
 # the answer — top it up with what the agent recorded along the way.
 _SHORT_REPORT_CHARS = 200

AutoCua/linux/controller/tool/browser_agent.py

@@ -64,6 +64,7 @@
     "perplexity": "PERPLEXITY_API_KEY",
     "together": "TOGETHER_API_KEY",
     "cerebras": "CEREBRAS_API_KEY",
+    "requesty": "REQUESTY_API_KEY",
     "aws": "AWS_API_KEY",
 }
 # A `done` value shorter than this is almost certainly an activity log, so it

AutoCua/linux/controller/tool/web/service.py

@@ -14,7 +14,7 @@
 # Providers with no native web-search API: the query goes to the browser
 # agent on the same provider+model instead (see web_agent.py). Explicit
 # membership, not a catch-all — a typo'd provider must still fail fast below.
-_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "aws"})
+_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "requesty", "aws"})
 
 
 class WebService:

AutoCua/linux/controller/tool/web/web_agent.py

@@ -48,7 +48,7 @@
 # __main__ has no --api_key flag, and its .env loader never overrides a var
 # that is already set.
 _PROVIDER_ENV_KEY = {"together": "TOGETHER_API_KEY", "cerebras": "CEREBRAS_API_KEY",
-                     "aws": "AWS_API_KEY"}
+                     "requesty": "REQUESTY_API_KEY", "aws": "AWS_API_KEY"}
 # A `done` value shorter than this is almost certainly an activity log, not
 # the answer — top it up with what the agent recorded along the way.
 _SHORT_REPORT_CHARS = 200

AutoCua/linux/remote_connection/telegram/service.py

@@ -78,6 +78,7 @@
     "perplexity": "PERPLEXITY_API_KEY",
     "together":   "TOGETHER_API_KEY",
     "cerebras":   "CEREBRAS_API_KEY",
+    "requesty":   "REQUESTY_API_KEY",
     "aws":        "AWS_API_KEY",
 }
 

AutoCua/llm_provider/__init__.py

@@ -72,6 +72,7 @@ def tls_verify():
     "perplexity": "https://api.perplexity.ai/",
     "together": "https://api.together.xyz/",
     "cerebras": "https://api.cerebras.ai/",
+    "requesty": "https://router.requesty.ai/",
 }
 
 

AutoCua/llm_provider/llm_manager.py

@@ -21,6 +21,7 @@
     "perplexity": "PerplexityProvider",
     "together": "TogetherProvider",
     "cerebras": "CerebrasProvider",
+    "requesty": "RequestyProvider",
     "aws": "AWSProvider",
 }
 
@@ -478,6 +479,14 @@ def build_provider(provider: str, model_short_name: str, model_info: dict, tools
         return _provider_class("cerebras")(api_key, cli_agent, model_info,
                                 tools=tools,
                                 max_tokens=max_tokens)
+    elif provider == "requesty":
+        # Priority: Runtime key > .env fallback
+        api_key = _key(api_key, 'REQUESTY_API_KEY')
+        if not api_key:
+            raise ValueError("Requesty API key not provided, not in .env, and not saved in the app's Settings")
+        return _provider_class("requesty")(api_key, cli_agent, model_info,
+                                tools=tools,
+                                max_tokens=max_tokens)
     elif provider == "aws":
         # Priority: Runtime key > .env fallback
         api_key = _key(api_key, 'AWS_API_KEY')

AutoCua/llm_provider/requesty/__init__.py

@@ -0,0 +1,4 @@
+from .service import RequestyProvider
+from .view import MODEL_MAPPINGS, get_model_info
+
+__all__ = ['RequestyProvider', 'MODEL_MAPPINGS', 'get_model_info']

AutoCua/llm_provider/requesty/service.py

@@ -0,0 +1,109 @@
+import json
+import requests
+from typing import Dict, Any, Optional
+from .. import LLM_HTTP_TIMEOUT, screenshots
+
+class RequestyProvider:
+    """Requesty API provider for LLM interactions (OpenAI-compatible chat completions)"""
+
+    def __init__(self, api_key: str, cli_agent: bool = False, model_info: dict = None, tools: list = None, max_tokens: int = None):
+        self.api_key = api_key
+        self.max_tokens = max_tokens
+        self.api_url = "https://router.requesty.ai/v1/chat/completions"
+        self.session = requests.Session()
+        self.cli_agent = cli_agent
+        self.model_info = model_info or {}
+        # Native tool calling: OpenAI-format function tools, the model's
+        # output contract. None only for mode="text" (plain prose).
+        self.tools = tools or None
+
+    def send_request(self, messages: list, model: str, annotated_screenshot_base64: Optional[str] = None) -> Dict[str, Any]:
+        """Send request to Requesty API"""
+
+        # If screenshot is provided and NOT cli_agent, modify the user message to include the annotated image
+        if annotated_screenshot_base64 and not self.cli_agent and len(messages) > 1:
+            user_message = messages[-1]["content"]
+            messages[-1]["content"] = [
+                {
+                    "type": "text",
+                    "text": user_message
+                },
+            ] + [
+                {
+                    "type": "image_url",
+                    "image_url": {
+                        "url": f"data:image/jpeg;base64,{shot}"
+                    }
+                }
+                for shot in screenshots(annotated_screenshot_base64)
+            ]
+
+        headers = {
+            "Authorization": f"Bearer {self.api_key}",
+            "Content-Type": "application/json"
+        }
+
+        data = {
+            "model": model,
+            "messages": messages,
+            "temperature": 0.2,
+            # Output cap. The handoff manager passes 10k (LLMManager.max_tokens);
+            # everything else keeps this default.
+            "max_tokens": self.max_tokens or 10000,
+        }
+
+        # No tools (mode="text") -> plain text: omit the tool params entirely.
+        if self.tools:
+            data["tools"] = self.tools
+            # "required" - every turn must call at least one tool; a text-only
+            # turn is never a valid step. Canonical rationale: see
+            # openrouter/service.py.
+            data["tool_choice"] = self.model_info.get("tool_choice") or "required"
+
+        try:
+            response = self.session.post(self.api_url, json=data, headers=headers, timeout=LLM_HTTP_TIMEOUT)
+            response.raise_for_status()
+            result = response.json()
+            if self.tools:
+                return _normalize_tool_response(result)
+            return result
+        except requests.exceptions.RequestException as e:
+            error_msg = f"Requesty API request failed: {str(e)}"
+            if hasattr(e, 'response') and e.response is not None:
+                error_msg += f"\nResponse: {e.response.text}"
+            raise Exception(error_msg)
+
+
+def _normalize_tool_response(result: dict) -> dict:
+    """Normalize an OpenAI-format tool-call response: keep the usual
+    choices/message shape but replace tool_calls with
+    [{"id", "name", "arguments": dict}] (arguments JSON-decoded). The `id` is
+    echoed back by the loop so each tool result matches its call."""
+    message = ((result.get("choices") or [{}])[0].get("message")) or {}
+    calls = []
+    for i, tc in enumerate(message.get("tool_calls") or []):
+        fn = (tc or {}).get("function") or {}
+        raw_args = fn.get("arguments")
+        if isinstance(raw_args, dict):
+            args = raw_args
+        else:
+            try:
+                args = json.loads(raw_args or "{}")
+            except Exception:
+                args = {}
+        if not isinstance(args, dict):
+            args = {}
+        calls.append({
+            "id": str((tc or {}).get("id") or "") or f"call_{i}",
+            "name": fn.get("name") or "",
+            "arguments": args,
+        })
+    return {
+        "choices": [{
+            "message": {
+                "content": message.get("content") or "",
+                "tool_calls": calls,
+            }
+        }],
+        "usage": result.get("usage", {}),
+    }

AutoCua/llm_provider/requesty/view.py

@@ -0,0 +1,40 @@
+# Model mappings for Requesty provider
+# Maps user-friendly names to actual API model names
+
+# Requesty is an OpenAI-compatible router: one key, `vendor/model` ids, the
+# same chat/completions shape for every backend. All entries below are
+# image+text in / text out with native (OpenAI-format) tool calling, checked
+# against https://router.requesty.ai/v1/models. Any other id from that list
+# can be typed as the model name and is sent verbatim.
+#
+# Reasoning is deliberately left unset, the same choice together/ and groq/
+# make: with no field on the request each model keeps its own default.
+# Temperature is handled once in service.py (0.2).
+MODEL_MAPPINGS = {
+    "claude-sonnet-5": {
+        "api_name": "anthropic/claude-sonnet-5",
+        "vision": True,
+        "display_name": "Claude Sonnet 5"
+    },
+    "gemini-3.6-flash": {
+        "api_name": "google/gemini-3.6-flash",
+        "vision": True,
+        "display_name": "Gemini 3.6 Flash"
+    },
+    "gpt-5.4": {
+        "api_name": "openai/gpt-5.4",
+        "vision": True,
+        "display_name": "GPT-5.4"
+    }
+}
+
+def get_model_info(short_name: str) -> dict:
+    """Get full model information from short name"""
+    if short_name in MODEL_MAPPINGS:
+        return MODEL_MAPPINGS[short_name]
+    # If not found, assume it's already a full model name
+    return {
+        "api_name": short_name,
+        "vision": True,
+        "display_name": short_name
+    }

AutoCua/mac/controller/tool/browser_agent.py

@@ -64,6 +64,7 @@
     "perplexity": "PERPLEXITY_API_KEY",
     "together": "TOGETHER_API_KEY",
     "cerebras": "CEREBRAS_API_KEY",
+    "requesty": "REQUESTY_API_KEY",
     "aws": "AWS_API_KEY",
 }
 # A `done` value shorter than this is almost certainly an activity log, so it

AutoCua/mac/controller/tool/web/service.py

@@ -14,7 +14,7 @@
 # Providers with no native web-search API: the query goes to the browser
 # agent on the same provider+model instead (see web_agent.py). Explicit
 # membership, not a catch-all — a typo'd provider must still fail fast below.
-_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "aws"})
+_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "requesty", "aws"})
 
 
 class WebService:

AutoCua/mac/controller/tool/web/web_agent.py

@@ -48,7 +48,7 @@
 # __main__ has no --api_key flag, and its .env loader never overrides a var
 # that is already set.
 _PROVIDER_ENV_KEY = {"together": "TOGETHER_API_KEY", "cerebras": "CEREBRAS_API_KEY",
-                     "aws": "AWS_API_KEY"}
+                     "requesty": "REQUESTY_API_KEY", "aws": "AWS_API_KEY"}
 # A `done` value shorter than this is almost certainly an activity log, not
 # the answer — top it up with what the agent recorded along the way.
 _SHORT_REPORT_CHARS = 200

AutoCua/mac/remote_connection/telegram/service.py

@@ -76,6 +76,7 @@
     "perplexity": "PERPLEXITY_API_KEY",
     "together":   "TOGETHER_API_KEY",
     "cerebras":   "CEREBRAS_API_KEY",
+    "requesty":   "REQUESTY_API_KEY",
     "aws":        "AWS_API_KEY",
 }
 

AutoCua/ui/index.tsx

@@ -79,6 +79,7 @@ const PROVIDERS: [string, string, string, string][] = [
     ['perplexity', 'https://docs.perplexity.ai/guides/getting-started', 'Perplexity', 'pplx-...'],
     ['together', 'https://api.together.xyz/settings/api-keys', 'Together AI', 'Enter API key...'],
     ['cerebras', 'https://cloud.cerebras.ai', 'Cerebras', 'csk-...'],
+    ['requesty', 'https://app.requesty.ai/api-keys', 'Requesty', 'Enter API key...'],
     ['aws', 'https://console.aws.amazon.com/bedrock/home#/api-keys', 'AWS', 'ABSK...'],
     ['google', 'https://aistudio.google.com/apikey', 'Google', 'AIza...'],
 ];

AutoCua/ui/script.js

@@ -33,6 +33,7 @@ document.addEventListener('DOMContentLoaded', () => {
         perplexity: null,
         together: null,
         cerebras: null,
+        requesty: null,
         aws: null
     };
 

AutoCua/ui/service.py

@@ -891,6 +891,10 @@ def get_llm_providers():
             cerebras_models = importlib.import_module(f"{base}.cerebras.view").MODEL_MAPPINGS
         except ModuleNotFoundError:
             cerebras_models = None
+        try:
+            requesty_models = importlib.import_module(f"{base}.requesty.view").MODEL_MAPPINGS
+        except ModuleNotFoundError:
+            requesty_models = None
         try:
             aws_models = importlib.import_module(f"{base}.aws.view").MODEL_MAPPINGS
         except ModuleNotFoundError:
@@ -914,6 +918,8 @@ def format_models(mappings):
             providers.append({'id': 'together', 'name': 'together', 'models': format_models(together_models)})
         if cerebras_models:
             providers.append({'id': 'cerebras', 'name': 'cerebras', 'models': format_models(cerebras_models)})
+        if requesty_models:
+            providers.append({'id': 'requesty', 'name': 'requesty', 'models': format_models(requesty_models)})
         if aws_models:
             providers.append({'id': 'aws', 'name': 'aws', 'models': format_models(aws_models)})
         return providers
@@ -931,6 +937,7 @@ def format_models(mappings):
     'perplexity': 'PERPLEXITY_API_KEY',
     'together': 'TOGETHER_API_KEY',
     'cerebras': 'CEREBRAS_API_KEY',
+    'requesty': 'REQUESTY_API_KEY',
     'aws': 'AWS_API_KEY',
 }
 

AutoCua/windows/controller/tool/browser_agent.py

@@ -64,6 +64,7 @@
     "perplexity": "PERPLEXITY_API_KEY",
     "together": "TOGETHER_API_KEY",
     "cerebras": "CEREBRAS_API_KEY",
+    "requesty": "REQUESTY_API_KEY",
     "aws": "AWS_API_KEY",
 }
 # A `done` value shorter than this is almost certainly an activity log, so it

AutoCua/windows/controller/tool/web/service.py

@@ -14,7 +14,7 @@
 # Providers with no native web-search API: the query goes to the browser
 # agent on the same provider+model instead (see web_agent.py). Explicit
 # membership, not a catch-all - a typo'd provider must still fail fast below.
-_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "aws"})
+_BROWSER_FALLBACK_PROVIDERS = frozenset({"together", "cerebras", "requesty", "aws"})
 
 
 class WebService:

AutoCua/windows/controller/tool/web/web_agent.py

@@ -48,7 +48,7 @@
 # __main__ has no --api_key flag, and its .env loader never overrides a var
 # that is already set.
 _PROVIDER_ENV_KEY = {"together": "TOGETHER_API_KEY", "cerebras": "CEREBRAS_API_KEY",
-                     "aws": "AWS_API_KEY"}
+                     "requesty": "REQUESTY_API_KEY", "aws": "AWS_API_KEY"}
 # A `done` value shorter than this is almost certainly an activity log, not
 # the answer — top it up with what the agent recorded along the way.
 _SHORT_REPORT_CHARS = 200

AutoCua/windows/remote_connection/telegram/service.py

@@ -78,6 +78,7 @@
     "perplexity": "PERPLEXITY_API_KEY",
     "together":   "TOGETHER_API_KEY",
     "cerebras":   "CEREBRAS_API_KEY",
+    "requesty":   "REQUESTY_API_KEY",
     "aws":        "AWS_API_KEY",
 }