Feat: added support for non vision models STRIX_DISABLE_IMAGES flag

README.md

@@ -208,8 +208,14 @@ export LLM_API_KEY="your-api-key"
 # Optional
 export LLM_API_BASE="your-api-base-url"  # if using a local model, e.g. Ollama, LMStudio
 export PERPLEXITY_API_KEY="your-api-key"  # for search capabilities
+export STRIX_DISABLE_IMAGES="true"  # disable image/vision features for non-vision LLMs
 ```
 
+**Disabling Image Features**: If you're using an LLM that doesn't support images/vision (e.g., some local models), set `STRIX_DISABLE_IMAGES=true`. This will:
+- Disable browser automation tools (which rely on screenshots)
+- Skip adding screenshots to LLM messages
+- Improve performance and reduce costs for non-vision models
+
 [OpenAI's GPT-5](https://openai.com/api/) (`openai/gpt-5`) and [Anthropic's Claude Sonnet 4.5](https://claude.com/platform/api) (`anthropic/claude-sonnet-4-5`) are the recommended models for best results with Strix. We also support many [other options](https://docs.litellm.ai/docs/providers), including cloud and local models, though their performance and reliability may vary.
 
 ## 🤝 Contributing

strix/llm/config.py

@@ -8,6 +8,7 @@ def __init__(
         enable_prompt_caching: bool = True,
         prompt_modules: list[str] | None = None,
         timeout: int | None = None,
+        disable_images: bool | None = None,
     ):
         self.model_name = model_name or os.getenv("STRIX_LLM", "openai/gpt-5")
 
@@ -17,4 +18,9 @@ def __init__(
         self.enable_prompt_caching = enable_prompt_caching
         self.prompt_modules = prompt_modules or []
 
-        self.timeout = timeout or int(os.getenv("LLM_TIMEOUT", "300"))
+        self.timeout = timeout or int(os.getenv("LLM_TIMEOUT", "600"))
+
+        if disable_images is None:
+            self.disable_images = os.getenv("STRIX_DISABLE_IMAGES", "false").lower() == "true"
+        else:
+            self.disable_images = disable_images

strix/llm/memory_compressor.py

@@ -4,6 +4,8 @@
 
 import litellm
 
+from strix.tools import is_images_disabled
+
 
 logger = logging.getLogger(__name__)
 
@@ -125,6 +127,17 @@ def _summarize_messages(
 
 
 def _handle_images(messages: list[dict[str, Any]], max_images: int) -> None:
+    if is_images_disabled():
+        for msg in messages:
+            content = msg.get("content", [])
+            if isinstance(content, list):
+                msg["content"] = [
+                    item
+                    for item in content
+                    if not (isinstance(item, dict) and item.get("type") == "image_url")
+                ]
+        return
+
     image_count = 0
     for msg in reversed(messages):
         content = msg.get("content", [])

strix/tools/__init__.py

@@ -24,9 +24,16 @@
 
 HAS_PERPLEXITY_API = bool(os.getenv("PERPLEXITY_API_KEY"))
 
+
+def is_images_disabled() -> bool:
+    return os.getenv("STRIX_DISABLE_IMAGES", "false").lower() == "true"
+
+
 if not SANDBOX_MODE:
     from .agents_graph import *  # noqa: F403
-    from .browser import *  # noqa: F403
+
+    if not is_images_disabled():
+        from .browser import *  # noqa: F403
     from .file_edit import *  # noqa: F403
     from .finish import *  # noqa: F403
     from .notes import *  # noqa: F403
@@ -39,7 +46,8 @@
     if HAS_PERPLEXITY_API:
         from .web_search import *  # noqa: F403
 else:
-    from .browser import *  # noqa: F403
+    if not is_images_disabled():
+        from .browser import *  # noqa: F403
     from .file_edit import *  # noqa: F403
     from .notes import *  # noqa: F403
     from .proxy import *  # noqa: F403
@@ -55,6 +63,7 @@
     "get_tool_by_name",
     "get_tool_names",
     "get_tools_prompt",
+    "is_images_disabled",
     "needs_agent_state",
     "process_tool_invocations",
     "register_tool",

strix/tools/executor.py

@@ -177,17 +177,22 @@ def _update_tracer_with_result(
 def _format_tool_result(tool_name: str, result: Any) -> tuple[str, list[dict[str, Any]]]:
     images: list[dict[str, Any]] = []
 
-    screenshot_data = extract_screenshot_from_result(result)
-    if screenshot_data:
-        images.append(
-            {
-                "type": "image_url",
-                "image_url": {"url": f"data:image/png;base64,{screenshot_data}"},
-            }
-        )
-        result_str = remove_screenshot_from_result(result)
+    from strix.tools import is_images_disabled
+
+    if not is_images_disabled():
+        screenshot_data = extract_screenshot_from_result(result)
+        if screenshot_data:
+            images.append(
+                {
+                    "type": "image_url",
+                    "image_url": {"url": f"data:image/png;base64,{screenshot_data}"},
+                }
+            )
+            result_str = remove_screenshot_from_result(result)
+        else:
+            result_str = result
     else:
-        result_str = result
+        result_str = remove_screenshot_from_result(result) if isinstance(result, dict) else result
 
     if result_str is None:
         final_result_str = f"Tool {tool_name} executed successfully"

strix/tools/registry.py

@@ -169,8 +169,18 @@ def should_execute_in_sandbox(tool_name: str) -> bool:
 
 
 def get_tools_prompt() -> str:
+    try:
+        from strix.tools import is_images_disabled
+
+        images_disabled = is_images_disabled()
+    except ImportError:
+        images_disabled = False
+
     tools_by_module: dict[str, list[dict[str, Any]]] = {}
     for tool in tools:
+        if images_disabled and tool.get("module") == "browser":
+            continue
+
         module = tool.get("module", "unknown")
         if module not in tools_by_module:
             tools_by_module[module] = []