[docs] AgentBase — Built-in Tools MCP header + architecture diagrams + Runtime Outbound IP — VI + EN

English/ai-stack/agent-base/agent-runtime/manage-runtime.md

@@ -33,6 +33,20 @@ While the Runtime is in `STOPPED` state, all requests to its Endpoint will retur
 
 ***
 
+## View a Runtime's Outbound IP
+
+The Runtime detail page shows the **Outbound IP** — the address the Runtime uses when calling out to your infrastructure (internal APIs, databases, MCP Servers in your VPC).
+
+**Step 1:** Open **Agent Runtime** → click the Runtime name to open its detail page
+
+**Step 2:** Check the **Outbound IP** field on the detail page
+
+{% hint style="info" %}
+For **Private** Runtimes: if a server or VM in your VPC uses an IP-based firewall / whitelist, add the Runtime's **Outbound IP** to the allow list — otherwise, calls from the Runtime to your infrastructure will be blocked.
+{% endhint %}
+
+***
+
 ## Update Runtime / Deploy a New Version
 
 Each update creates a new immutable **Version**. The DEFAULT Endpoint automatically points to the latest version — the old version continues serving traffic until the deployment completes.

English/ai-stack/agent-base/built-in-tools/README.md

@@ -19,24 +19,11 @@ Built-in Tools is currently being rolled out by **whitelist**. If your account i
 
 ## Architecture
 
-Each of your sessions runs inside its own microVM with a dedicated kernel, claimed from a pre-warmed VM pool, so creating a session takes only a few seconds.
-
-```
-Agent / MCP client
-        │  MCP (HTTPS)
-        ▼
-Built-in Tools gateway
-  (Code Interpreter · Browser)
-        │
-        ▼
-AgentBase Runtime ── OpenSandbox (warm VM pool)
-        │
-        ▼
-Dedicated vKS node group (nested virtualization)
-        │  Kata Containers + Cloud Hypervisor
-        ▼
-Each session = 1 microVM with its own kernel
-```
+Each of your sessions runs inside its own microVM with a dedicated kernel, claimed from a pre-warmed VM pool, so creating a session takes only a few seconds. Both tools share the same sandbox foundation — they differ in the product layer on top:
+
+<figure><img src="../../../.gitbook/assets/Agentbase-image/agentbase-code-interpreter-architecture.svg" alt=""><figcaption><p>Code Interpreter — the agent calls tools over MCP, code runs in an isolated sandbox microVM</p></figcaption></figure>
+
+<figure><img src="../../../.gitbook/assets/Agentbase-image/agentbase-browser-architecture.svg" alt=""><figcaption><p>Browser — the agent drives Chromium inside a sandbox microVM over CDP</p></figcaption></figure>
 
 Main layers:
 

English/ai-stack/agent-base/built-in-tools/use-browser.md

@@ -64,6 +64,10 @@ Recording ─ saved when the session is deleted ─────┘
 - **Live view** — a human watches Chrome's screen live over noVNC, embedded in the portal via an iframe.
 - **Recording** — a session recording is preserved when the session ends.
 
+The full flow: your agent (with an automation library such as Playwright or browser-use) receives a query → the LLM picks tool calls (`click`, `type`, `goto`...) → CDP commands travel over the automation stream into the Chromium inside the session → page content and screenshots return to the agent:
+
+<figure><img src="../../../.gitbook/assets/Agentbase-image/agentbase-browser-architecture.svg" alt=""><figcaption><p>Browser architecture — the agent drives Chromium inside a sandbox microVM over CDP</p></figcaption></figure>
+
 Browser follows the same pattern as [Code Interpreter](use-code-interpreter.md): create the resource, wait for `ACTIVE`, create sessions (one VM per session), delete when done.
 
 ---

English/ai-stack/agent-base/built-in-tools/use-code-interpreter.md

@@ -52,6 +52,16 @@ Tokens from a **Root user, IAM user, or Service Account** all work. If your agen
 
 ---
 
+## How it works
+
+Your agent receives a user query, the LLM picks a tool call, and the tool call travels over the MCP connection into the Code Interpreter — running inside its own sandbox microVM and returning the result to the agent:
+
+<figure><img src="../../../.gitbook/assets/Agentbase-image/agentbase-code-interpreter-architecture.svg" alt=""><figcaption><p>Code Interpreter architecture — the agent calls tools over MCP, code runs in an isolated sandbox microVM</p></figcaption></figure>
+
+A Code Interpreter is **created once and reused** across sessions — each session is still its own microVM.
+
+---
+
 ## Usage lifecycle
 
 One round of using Code Interpreter goes through these steps:

English/ai-stack/agent-base/private-networking.md

@@ -90,6 +90,16 @@ Fill in the remaining fields (Inbound Auth, MCP Servers, Policy Group...) and cl
 
 ***
 
+## Whitelist a Private Runtime's Outbound IP
+
+When a Private Runtime calls into your infrastructure (databases, internal APIs, MCP Servers), the request originates from the Runtime's **Outbound IP**. If your side filters by IP (security group, firewall, whitelist), you need to allow this IP:
+
+1. Open the Runtime detail page — see [Manage Runtime](agent-runtime/manage-runtime.md#view-a-runtimes-outbound-ip)
+2. Copy the **Outbound IP** shown on the detail page
+3. Add the IP to the whitelist / security group / firewall of the server or VM in your VPC
+
+***
+
 ## Result
 
 After creation with Private mode: