diff --git a/ambient/agent.py b/ambient/agent.py
index 0113473..743774d 100644
--- a/ambient/agent.py
+++ b/ambient/agent.py
@@ -354,8 +354,13 @@ async def run_agent(
             print(f"Turn {turn}")
             turn_trajectory: dict = {"turn": turn, "messages": list(messages)}
 
+            # Agent (orchestrator) endpoint may differ from the vision endpoint:
+            # vision tools use llm_base_url (e.g. local vLLM); the agent uses agent_base_url
+            # (e.g. OpenRouter/deepseek) when set, falling back to the llm_* pair.
+            _agent_base = settings.agent_base_url or settings.llm_base_url
+            _agent_key = settings.agent_api_key or settings.llm_api_key
             resp = await _send_request(
-                settings.llm_base_url, settings.llm_api_key, model, messages
+                _agent_base, _agent_key, model, messages
             )
             if "content" not in resp:
                 print(f"No content in response: {resp}")
diff --git a/ambient/llm.py b/ambient/llm.py
index 68e356a..ebd2d5c 100644
--- a/ambient/llm.py
+++ b/ambient/llm.py
@@ -28,6 +28,22 @@ def video_to_data_url(path: str, mime="video/mp4") -> str:
     return f"data:{mime};base64,{b64}"
 
 
+def _clip_to_frame_urls(mp4: str, fps: float = 0.5, cap: int = 28, dim: int = 768) -> list:
+    """Sample chronological frames from a clip mp4 as base64 image data-URLs.
+    Used for image-only vision models (e.g. qwen) that reject video_url input."""
+    import subprocess, tempfile, glob as _glob
+    d = tempfile.mkdtemp(prefix="clipfr_")
+    subprocess.run(
+        ["ffmpeg", "-nostdin", "-loglevel", "error", "-i", mp4, "-vf",
+         f"fps={fps},scale='if(gt(iw,ih),{dim},-2)':'if(gt(iw,ih),-2,{dim})'", "-q:v", "4",
+         os.path.join(d, "f%04d.jpg")], check=False)
+    out = []
+    for p in sorted(_glob.glob(os.path.join(d, "*.jpg")))[:cap]:
+        with open(p, "rb") as f:
+            out.append("data:image/jpeg;base64," + base64.b64encode(f.read()).decode())
+    return out
+
+
 def construct_payload(clips: Optional[List[Clip]] = None, frames: Optional[List[Frame]] = None):
     payload = []
     if clips is None and frames is None:
@@ -38,28 +54,26 @@ def construct_payload(clips: Optional[List[Clip]] = None, frames: Optional[List[
             is_clip_url = clip.clip_url is not None
             is_clip_file_path = clip.clip_file_path is not None and os.path.exists(clip.clip_file_path)
 
-            if is_clip_url:
-                url = clip.clip_url
-            elif is_clip_file_path:
-                url = video_to_data_url(clip.clip_file_path)
-            else:
-                raise ValueError(f"Clip {clip.id} has no valid url or file path")
-
-
             timestamp = f"Timestamp: {clip.start_time} seconds to {clip.end_time} seconds"
-
-            payload.extend(
-                [
-                    {
-                        "type": "text",
-                        "text": f"Clip ID: {clip.id}\n{timestamp}",
-                    },
-                    {
-                        "type": "video_url",
-                        "video_url": {"url": url},
-                    },
-                ]
-            )
+            image_only = "gemini" not in (settings.llm_model or "").lower()
+
+            if is_clip_file_path and image_only:
+                # image-only vision model (e.g. qwen3.6-27b): send sampled frames, not video_url
+                payload.append({"type": "text",
+                                "text": f"Clip ID: {clip.id}\n{timestamp}\n(chronological sampled frames)"})
+                for u in _clip_to_frame_urls(clip.clip_file_path):
+                    payload.append({"type": "image_url", "image_url": {"url": u}})
+            else:
+                if is_clip_url:
+                    url = clip.clip_url
+                elif is_clip_file_path:
+                    url = video_to_data_url(clip.clip_file_path)
+                else:
+                    raise ValueError(f"Clip {clip.id} has no valid url or file path")
+                payload.extend([
+                    {"type": "text", "text": f"Clip ID: {clip.id}\n{timestamp}"},
+                    {"type": "video_url", "video_url": {"url": url}},
+                ])
 
     if frames is not None:
         start_time = frames[0].timestamp