diff --git a/assets/css/markdown.css b/assets/css/markdown.css
index fe50705..024f539 100644
--- a/assets/css/markdown.css
+++ b/assets/css/markdown.css
@@ -92,7 +92,7 @@
 }
 
 .markdown table {
-  @apply w-full border-collapse my-4 text-sm;
+  @apply block w-full border-collapse my-4 text-sm overflow-x-auto;
 }
 .markdown thead {
   @apply bg-cyan-950;
diff --git a/assets/js/app.js b/assets/js/app.js
index 7b114fd..e93ef47 100644
--- a/assets/js/app.js
+++ b/assets/js/app.js
@@ -24,6 +24,52 @@ import topbar from "../vendor/topbar"
 
 let Hooks = {}
 
+// Renders a complete markdown string client-side on mount.
+// The raw markdown is passed as the data-md attribute.
+Hooks.MarkdownRender = {
+  mounted() {
+    const smd = window.smd
+    const content = this.el.dataset.md
+    if (!content) return
+    const parser = smd.parser(smd.default_renderer(this.el))
+    smd.parser_write(parser, content)
+    smd.parser_end(parser)
+  }
+}
+
+// Streams markdown chunks into the element using the streaming-markdown parser.
+// The server sends push_event(socket, eventName, %{chunk: "..."}) for each chunk.
+// data-event on the element controls which event this hook listens for.
+// The server sends push_event(socket, eventName, %{chunk: "..."}) for each chunk.
+// data-event on the element controls which event this hook listens for.
+Hooks.MarkdownStream = {
+  mounted() {
+    const smd = window.smd
+    const DOMPurify = window.DOMPurify
+    this._chunks = ""
+    this._parser = smd.parser(smd.default_renderer(this.el))
+    const eventName = this.el.dataset.event
+    this.handleEvent(eventName, ({chunk}) => {
+      this._chunks += chunk
+      // Sanitize all accumulated chunks to detect injection attacks.
+      DOMPurify.sanitize(this._chunks)
+      if (DOMPurify.removed.length > 0) {
+        // Insecure content detected — stop rendering immediately.
+        smd.parser_end(this._parser)
+        this._parser = null
+        return
+      }
+      if (this._parser) smd.parser_write(this._parser, chunk)
+    })
+  },
+  destroyed() {
+    if (this._parser) {
+      window.smd.parser_end(this._parser)
+      this._parser = null
+    }
+  }
+}
+
 Hooks.ScrollBottom = {
   mounted() {
     this.scrollToBottom()
diff --git a/lib/elixir_ai_web/components/chat_message.ex b/lib/elixir_ai_web/components/chat_message.ex
index 5591788..b4f9eb0 100644
--- a/lib/elixir_ai_web/components/chat_message.ex
+++ b/lib/elixir_ai_web/components/chat_message.ex
@@ -1,6 +1,5 @@
 defmodule ElixirAiWeb.ChatMessage do
   use Phoenix.Component
-  alias ElixirAiWeb.Markdown
   alias Phoenix.LiveView.JS
 
   attr :content, :string, required: true
@@ -60,20 +59,62 @@ defmodule ElixirAiWeb.ChatMessage do
   attr :reasoning_content, :string, default: nil
   attr :tool_calls, :list, default: []
 
+  # Renders the in-progress streaming message. Content and reasoning are rendered
+  # entirely client-side via the MarkdownStream hook — the server sends push_event
+  # chunks instead of re-rendering the full markdown on every token.
   def streaming_assistant_message(assigns) do
-    assigns =
-      assigns
-      |> assign(:_reasoning_id, "reasoning-stream")
-      |> assign(:_expanded, true)
-
     ~H"""
-    <.message_bubble
-      reasoning_id={@_reasoning_id}
-      content={@content}
-      reasoning_content={@reasoning_content}
-      tool_calls={@tool_calls}
-      expanded={@_expanded}
-    />
+    <div class="mb-2 text-sm text-left">
+      <!-- Reasoning section — only shown once reasoning_content is non-empty.
+           The div is always in the DOM so the hook mounts before chunks arrive. -->
+      <div id="stream-reasoning-wrap">
+        <%= if @reasoning_content && @reasoning_content != "" do %>
+          <button
+            type="button"
+            class="flex items-center text-cyan-500/60 hover:text-cyan-300 transition-colors duration-150 cursor-pointer"
+            phx-click={
+              JS.toggle_class("collapsed", to: "#reasoning-stream")
+              |> JS.toggle_class("rotate-180", to: "#reasoning-stream-chevron")
+            }
+            aria-label="Toggle reasoning"
+          >
+            <div class="flex items-center gap-1 text-cyan-100/40 ps-2 mb-1">
+              <span class="text-xs">reasoning</span>
+              <svg
+                id="reasoning-stream-chevron"
+                xmlns="http://www.w3.org/2000/svg"
+                viewBox="0 0 20 20"
+                fill="currentColor"
+                class="w-3 h-3 transition-transform duration-300"
+              >
+                <path
+                  fill-rule="evenodd"
+                  d="M9.47 6.47a.75.75 0 0 1 1.06 0l4.25 4.25a.75.75 0 1 1-1.06 1.06L10 8.06l-3.72 3.72a.75.75 0 0 1-1.06-1.06l4.25-4.25Z"
+                  clip-rule="evenodd"
+                />
+              </svg>
+            </div>
+          </button>
+        <% end %>
+        <div
+          id="reasoning-stream"
+          phx-hook="MarkdownStream"
+          phx-update="ignore"
+          data-event="reasoning_chunk"
+          class="reasoning-content block px-3 py-2 rounded-lg bg-cyan-950/50 text-cyan-400 italic text-xs max-w-prose mb-1 markdown"
+        ></div>
+      </div>
+      <%= for tool_call <- @tool_calls do %>
+        <.tool_call_item tool_call={tool_call} />
+      <% end %>
+      <div
+        id="stream-content"
+        phx-hook="MarkdownStream"
+        phx-update="ignore"
+        data-event="md_chunk"
+        class="inline-block px-3 py-2 rounded-lg max-w-prose markdown bg-cyan-950/50"
+      ></div>
+    </div>
     """
   end
 
@@ -115,21 +156,26 @@ defmodule ElixirAiWeb.ChatMessage do
         </button>
         <div
           id={@reasoning_id}
+          phx-hook="MarkdownRender"
+          phx-update="ignore"
+          data-md={@reasoning_content}
           class={[
             "reasoning-content block px-3 py-2 rounded-lg bg-cyan-950/50 text-cyan-400 italic text-xs max-w-prose mb-1 markdown",
             !@expanded && "collapsed"
           ]}
-        >
-          {Markdown.render(@reasoning_content)}
-        </div>
+        ></div>
       <% end %>
       <%= for tool_call <- @tool_calls do %>
         <.tool_call_item tool_call={tool_call} />
       <% end %>
       <%= if @content && @content != "" do %>
-        <div class="inline-block px-3 py-2 rounded-lg max-w-prose markdown bg-cyan-950/50">
-          {Markdown.render(@content)}
-        </div>
+        <div
+          id={"#{@reasoning_id}-content"}
+          phx-hook="MarkdownRender"
+          phx-update="ignore"
+          data-md={@content}
+          class="inline-block px-3 py-2 rounded-lg max-w-prose markdown bg-cyan-950/50"
+        ></div>
       <% end %>
     </div>
     """
diff --git a/lib/elixir_ai_web/components/layouts/root.html.heex b/lib/elixir_ai_web/components/layouts/root.html.heex
index 87ec6ff..b74e100 100644
--- a/lib/elixir_ai_web/components/layouts/root.html.heex
+++ b/lib/elixir_ai_web/components/layouts/root.html.heex
@@ -10,6 +10,12 @@
     <link phx-track-static rel="stylesheet" href={~p"/assets/app.css"} />
     <script defer phx-track-static type="text/javascript" src={~p"/assets/app.js"}>
     </script>
+    <script type="module">
+      import * as smd from "https://cdn.jsdelivr.net/npm/streaming-markdown/smd.min.js"
+      import DOMPurify from "https://cdn.jsdelivr.net/npm/dompurify/+esm"
+      window.smd = smd
+      window.DOMPurify = DOMPurify
+    </script>
   </head>
   <body class="bg-cyan-900 text-cyan-50">
     {@inner_content}
diff --git a/lib/elixir_ai_web/components/markdown.ex b/lib/elixir_ai_web/components/markdown.ex
deleted file mode 100644
index bbe2095..0000000
--- a/lib/elixir_ai_web/components/markdown.ex
+++ /dev/null
@@ -1,23 +0,0 @@
-defmodule ElixirAiWeb.Markdown do
-  @doc """
-  Converts a markdown string to sanitized HTML, safe for raw rendering.
-  Falls back to escaped plain text if the markdown is incomplete or invalid.
-  """
-  def render(nil), do: Phoenix.HTML.raw("")
-  def render(""), do: Phoenix.HTML.raw("")
-
-  def render(markdown) do
-    case Earmark.as_html(markdown, breaks: true, compact_output: true) do
-      {:ok, html, _warnings} ->
-        html
-        |> HtmlSanitizeEx.markdown_html()
-        |> Phoenix.HTML.raw()
-
-      {:error, html, _errors} ->
-        # Partial/invalid markdown — use whatever HTML was produced, still sanitize
-        html
-        |> HtmlSanitizeEx.markdown_html()
-        |> Phoenix.HTML.raw()
-    end
-  end
-end
diff --git a/lib/elixir_ai_web/live/chat_live.ex b/lib/elixir_ai_web/live/chat_live.ex
index 2aebb61..d7e13f4 100644
--- a/lib/elixir_ai_web/live/chat_live.ex
+++ b/lib/elixir_ai_web/live/chat_live.ex
@@ -131,7 +131,11 @@ defmodule ElixirAiWeb.ChatLive do
           socket.assigns.streaming_response.reasoning_content <> reasoning_content
     }
 
-    {:noreply, assign(socket, streaming_response: updated_response)}
+    # Update assign (controls toggle button visibility) and stream chunk to hook.
+    {:noreply,
+     socket
+     |> assign(streaming_response: updated_response)
+     |> push_event("reasoning_chunk", %{chunk: reasoning_content})}
   end
 
   def handle_info(
@@ -148,7 +152,11 @@ defmodule ElixirAiWeb.ChatLive do
       | content: socket.assigns.streaming_response.content <> text_content
     }
 
-    {:noreply, assign(socket, streaming_response: updated_response)}
+    # Update assign (accumulated for final message) and stream chunk to hook.
+    {:noreply,
+     socket
+     |> assign(streaming_response: updated_response)
+     |> push_event("md_chunk", %{chunk: text_content})}
   end
 
   def handle_info(:tool_calls_finished, socket) do
diff --git a/mix.exs b/mix.exs
index 4003c81..b901835 100644
--- a/mix.exs
+++ b/mix.exs
@@ -49,7 +49,6 @@ defmodule ElixirAi.MixProject do
        depth: 1},
       {:req, "~> 0.5"},
       {:html_sanitize_ex, "~> 1.4"},
-      {:earmark, "~> 1.4"},
       {:dotenvy, "~> 1.1.1"},
       {:telemetry_metrics, "~> 1.0"},
       {:telemetry_poller, "~> 1.0"},