diff --git a/examples/http_agent_server.py b/examples/http_agent_server.py
new file mode 100644
index 0000000..dc8de06
--- /dev/null
+++ b/examples/http_agent_server.py
@@ -0,0 +1,134 @@
+"""Example Flask server implementing the HttpAgent protocol.
+
+This is a minimal reference implementation showing the request/response
+contract for ``openadapt_evals.agents.HttpAgent``. Copy and adapt for
+your own model.
+
+Usage:
+    pip install flask pillow
+    python examples/http_agent_server.py
+
+    # Then run eval against it:
+    openadapt-evals run --agent http --agent-endpoint http://localhost:8080
+
+Protocol:
+    POST /act    - Receive observation, return action
+    POST /reset  - (Optional) Reset agent state between episodes
+    GET  /health - Health check (return 200)
+"""
+
+import base64
+import io
+import json
+import logging
+
+from flask import Flask, jsonify, request
+
+app = Flask(__name__)
+log = logging.getLogger(__name__)
+
+# ---------------------------------------------------------------------------
+# Replace this with your model loading and inference logic
+# ---------------------------------------------------------------------------
+
+
+def load_model():
+    """Load your model here. Called once at startup."""
+    log.info("Loading model... (replace this with your model loading code)")
+    # Example:
+    #   from transformers import AutoModelForVision2Seq, AutoProcessor
+    #   model = AutoModelForVision2Seq.from_pretrained("your-model")
+    #   processor = AutoProcessor.from_pretrained("your-model")
+    #   return model, processor
+    return None
+
+
+def predict(screenshot_bytes, instruction, viewport, step_count):
+    """Run your model on a screenshot and return an action dict.
+
+    Args:
+        screenshot_bytes: PNG image bytes (or None).
+        instruction: Task instruction string.
+        viewport: [width, height] or None.
+        step_count: How many steps have been taken so far.
+
+    Returns:
+        Action dict, e.g.:
+            {"type": "click", "x": 0.5, "y": 0.3}
+            {"type": "type", "text": "hello world"}
+            {"type": "key", "key": "Enter", "modifiers": ["ctrl"]}
+            {"type": "scroll", "scroll_direction": "down"}
+            {"type": "done"}
+
+        Coordinates should be in [0, 1] normalized range.
+    """
+    # --- Replace everything below with your inference code ---
+    log.info("Step %d: %s", step_count, instruction[:80])
+
+    # Dummy: always click center of screen
+    return {"type": "click", "x": 0.5, "y": 0.5}
+
+
+# ---------------------------------------------------------------------------
+# HTTP endpoints (you probably don't need to modify these)
+# ---------------------------------------------------------------------------
+
+MODEL = None
+
+
+@app.route("/act", methods=["POST"])
+def act():
+    """Receive observation, return action.
+
+    Request JSON:
+        screenshot_b64: str | null  - Base64-encoded PNG
+        instruction: str            - Task instruction
+        task_id: str                - Task identifier
+        viewport: [int, int] | null - [width, height]
+        accessibility_tree: dict | null
+        step_count: int             - Steps taken so far
+
+    Response JSON:
+        type: str           - "click", "type", "key", "scroll", "drag", "done"
+        x: float | null     - Normalized [0,1] x coordinate
+        y: float | null     - Normalized [0,1] y coordinate
+        text: str | null    - Text to type (for "type" action)
+        key: str | null     - Key name (for "key" action)
+        modifiers: list | null - ["ctrl", "shift", "alt"]
+        scroll_direction: str | null - "up", "down"
+        target_node_id: str | null   - A11y element ID
+    """
+    data = request.get_json(force=True)
+
+    # Decode screenshot
+    screenshot_bytes = None
+    if data.get("screenshot_b64"):
+        screenshot_bytes = base64.b64decode(data["screenshot_b64"])
+
+    action = predict(
+        screenshot_bytes=screenshot_bytes,
+        instruction=data.get("instruction", ""),
+        viewport=data.get("viewport"),
+        step_count=data.get("step_count", 0),
+    )
+
+    return jsonify(action)
+
+
+@app.route("/reset", methods=["POST"])
+def reset():
+    """Optional: reset agent state between episodes."""
+    log.info("Agent reset")
+    return jsonify({"status": "ok"})
+
+
+@app.route("/health", methods=["GET"])
+def health():
+    """Health check."""
+    return jsonify({"status": "ok"})
+
+
+if __name__ == "__main__":
+    logging.basicConfig(level=logging.INFO)
+    MODEL = load_model()
+    app.run(host="0.0.0.0", port=8080)