diff --git a/navi/api/deps.py b/navi/api/deps.py
index 57bf3b4..02df568 100644
--- a/navi/api/deps.py
+++ b/navi/api/deps.py
@@ -99,8 +99,25 @@
 
 
 async def register_mcp_tools(registry: ToolRegistry, manager: McpManager) -> None:
-    """(kept for backward compat; MCP tools are already registered at startup)."""
-    pass
+    """Register all currently connected MCP tools into the tool registry."""
+    from navi.mcp.tools import McpTool
+
+    for name, client in manager.clients.items():
+        if not client.connected:
+            continue
+        try:
+            tools = await client.list_tools()
+            for tool in tools:
+                mcp_tool = McpTool(
+                    server_name=name,
+                    tool_name=tool.name,
+                    description=tool.description or "",
+                    parameters=tool.inputSchema,
+                    manager=manager,
+                )
+                registry.register_external(mcp_tool)
+        except Exception as exc:
+            log.warning("Failed to register MCP tools for %r: %s", name, exc)
 
 
 # Backward-compat re-exports so nothing outside breaks
diff --git a/navi/main.py b/navi/main.py
index 16cfd13..7bd3ac1 100644
--- a/navi/main.py
+++ b/navi/main.py
@@ -25,6 +25,9 @@
     ),
 )
 
+# Suppress noisy MCP SDK health-check chatter.
+logging.getLogger("mcp.server.lowlevel.server").setLevel(logging.WARNING)
+
 _base = Path(__file__).parent.parent
 
 
diff --git a/navi/mcp/client.py b/navi/mcp/client.py
index 9807441..7d10f30 100644
--- a/navi/mcp/client.py
+++ b/navi/mcp/client.py
@@ -58,6 +58,10 @@
         if self._connected:
             return
 
+        # Ensure any stale transport/session from a prior disconnect or
+        # mark_disconnected() is fully torn down before we open fresh ones.
+        await self._cleanup()
+
         try:
             if self.config.is_stdio:
                 if not self.config.command:
@@ -134,6 +138,7 @@
         Used by the health-check loop when a server drops silently.
         """
         self._connected = False
+        self._session = None
 
     async def _cleanup(self) -> None:
         try:
diff --git a/navi/mcp/manager.py b/navi/mcp/manager.py
index 65cff7f..60ba80f 100644
--- a/navi/mcp/manager.py
+++ b/navi/mcp/manager.py
@@ -35,6 +35,10 @@
         self._health_check_task: asyncio.Task | None = None
         self._health_check_interval: float = 30.0
 
+        # Last known connected status per server — used to suppress duplicate
+        # "connected" toast notifications on every health-check poll.
+        self._connected_status: dict[str, bool] = {}
+
     @property
     def clients(self) -> dict[str, McpClient]:
         return self._clients
@@ -84,12 +88,14 @@
             try:
                 await client.connect()
                 self._clients[name] = client
+                self._connected_status[name] = True
                 if self._on_server_connected:
                     await self._on_server_connected(name)
             except Exception as exc:
                 logger.warning("MCP server %r failed to connect: %s", name, exc)
                 # Keep the client in the pool so health-check can retry later
                 self._clients[name] = client
+                self._connected_status[name] = False
 
     async def reload_all(self) -> None:
         """Re-read the config file and reconnect every server."""
@@ -109,6 +115,7 @@
             except Exception as exc:
                 logger.warning("MCP server %r disconnect error: %s", name, exc)
         self._clients.clear()
+        self._connected_status.clear()
 
     async def get_all_tools(self) -> list[tuple[str, Any]]:
         """Return ``(server_name, mcp_tool)`` for every tool on every server.
@@ -200,24 +207,31 @@
                 try:
                     await client.connect()
                     logger.info("MCP server %r reconnected by health check", name)
+                    self._connected_status[name] = True
                     if self._on_server_connected:
                         await self._on_server_connected(name)
                 except Exception as exc:
                     logger.debug("MCP health-check reconnect failed for %r: %s", name, exc)
                     continue
-            else:
-                # Live server — make sure it still responds
-                try:
-                    await client.list_tools()
-                except Exception:
-                    logger.warning("MCP server %r dropped during health check", name)
-                    client.mark_disconnected()
-                    await get_event_bus().publish(
-                        McpStatusUpdate(server_name=name, status="disconnected")
-                    )
-                    continue
+                continue
 
-            # If we get here the server is connected (either newly or verified)
+            # Live server — make sure it still responds
+            try:
+                await client.list_tools()
+            except Exception:
+                logger.warning("MCP server %r dropped during health check", name)
+                client.mark_disconnected()
+                self._connected_status[name] = False
+                await get_event_bus().publish(
+                    McpStatusUpdate(server_name=name, status="disconnected")
+                )
+                continue
+
+            # Server is still connected — only notify if it was previously
+            # known as disconnected (recovery), not on every routine poll.
+            if self._connected_status.get(name):
+                continue
+            self._connected_status[name] = True
             try:
                 tools = await client.list_tools()
                 await get_event_bus().publish(