Use the Agents window to build in an agent-first way.

Visual Studio Code 1.121

Follow us on LinkedIn, X, Bluesky

Release date: May 20, 2026

Downloads: Windows: x64 Arm64 | Mac: Universal Intel silicon | Linux: deb rpm tarball Arm snap

Welcome to the 1.121 release of Visual Studio Code. This release adds built-in Mermaid and HTML previews, streamlines terminal tool behavior for agents, and lets you run agent sessions on remote machines.

  • Remote agents: Monitor and control agent sessions on a remote machine from the Agents window.

  • Model configurability: Configure which models handle lightweight tasks like generating commit messages, titles, and more.

  • Mermaid diagram preview: Render Mermaid diagrams directly in the Markdown preview and notebooks.

  • HTML file preview: Preview local HTML files in the Integrated Browser without installing an extension.

  • Terminal tool optimizations: Consume less resources and tokens with more output compression and background terminal cleanup.

Happy Coding!

Agents

Agents Window (Preview)

We continue improvement to the Agents window, which is the agent-driven companion window brought as a preview to VS Code Stable in our last release.

You can open the Agents window in several ways, including the Open in Agents button in the VS Code title bar. To learn more about how it works and what you can do with it, visit the Agents window documentation.

Your feedback continues to be a great help in shaping Agents. If you've already been using it and providing feedback, thank you! Please continue to file issues on GitHub or browse existing issues.

We're also continuing to work on the broader extension story in the Agents window, including what extension enablement unlocks and how various extensions should behave in this environment. Whether you'd like to ideate on new scenarios that take advantage of running agents across projects, or share feedback on how your existing extension behaves in the Agents window, we'd love to collaborate with you through GitHub issues.

Remote agents (Preview)

The Agents window has experimental support for running agent sessions on a remote machine that you own and can connect to via SSH or dev tunnels. Learn more about remote agent sessions in our documentation.

Screenshot showing the Agents window remote tab for connecting to a remote machine.

Connecting to a remote

You can connect the Agents window to a remote machine in two ways:

  • SSH: pick from your existing ~/.ssh/config entries, or type a user@host.
  • Dev Tunnels: pick from tunnels you've already created by running code tunnel on the target machine.

How it works

This feature is similar to, but not the same as, VS Code's remote development extensions. The Agents window connects to the remote, and either downloads and installs the VS Code CLI (SSH) or connects to the running CLI server via a dev tunnel that you started. It starts a lightweight process called the "agent host", which hosts a new agent loop built on the Copilot SDK.

An important point to note is that the remote agent host is a long-lived process. Running sessions continue to run on the remote even if your client disconnects, so you can close your laptop while the remote agent continues working.

Agent Host Protocol

The connection between the Agents window and the agent host is a new open protocol called the Agent Host Protocol (AHP). We're developing it in the open as a standalone spec.

The key design principle of AHP is that it enables coordination of agent sessions across multiple clients simultaneously. This is how it differs from other protocols like ACP. An agent host manages authoritative state, synchronizes it to every connected client, and sequences all mutations through pure reducers.

Because AHP is an open protocol, anyone can build a client that connects to the VS Code CLI's agent host, or build an AHP agent host that VS Code can connect to.

Agents observability with OpenTelemetry and Grafana

In collaboration with the Azure Managed Grafana team, there is now a prebuilt Azure Managed Grafana dashboard for the OpenTelemetry signals that agents in VS Code emit. Point VS Code at an OTel Collector that forwards to Azure Application Insights, then import the Azure Managed Grafana dashboard to visualize agent operations, token usage, chat sessions, tool calls, and per-model response time and time to first token (TTFT).

See Monitor AI coding agents with Grafana for the end-to-end setup, and Monitor agent usage with OpenTelemetry for enabling export from VS Code.

Screenshot showing the GitHub Copilot Grafana dashboard with panels for operations, tokens, chat sessions, tool calls, and per-model latency.

Claude agent Auto permission mode (Preview)

Setting: github.copilot.chat.claudeAgent.allowAutoPermissions Open in VS Code Open in VS Code Insiders

The Claude Agent now supports Auto mode, which lets Claude execute without permission prompts. A separate classifier request reviews actions before they run, blocking anything that escalates beyond your request, targets unrecognized infrastructure, or appears driven by hostile content Claude read. This is useful for long-running tasks where you want to reduce prompt fatigue while still keeping background safety checks in place.

Claude agent auto mode

To see the Auto option in the permission mode picker, enable github.copilot.chat.claudeAgent.allowAutoPermissions Open in VS Code Open in VS Code Insiders .

Note: If you want fully unattended execution with no safety checks ("YOLO mode"), enable github.copilot.chat.claudeAgent.allowDangerouslySkipPermissions Open in VS Code Open in VS Code Insiders to allow "Bypass all permissions" to show up.

Language Models

This release includes several improvements to how you configure and manage language models in VS Code, giving you more control over which models you use for different tasks within VS Code. Learn more about language models in our documentation.

Configure utility models

Settings: chat.utilityModel Open in VS Code Open in VS Code Insiders , chat.utilitySmallModel Open in VS Code Open in VS Code Insiders

VS Code uses utility models in the background for chat-related tasks such as generating titles, summaries, commit messages, rename suggestions, prompt categorization, and intent detection. By default, these tasks use utility models provided by GitHub Copilot.

You can use your own available models, including Bring Your Own Key (BYOK) models, for these flows:

  • chat.utilityModel Open in VS Code Open in VS Code Insiders : Override the model used for general utility flows.
  • chat.utilitySmallModel Open in VS Code Open in VS Code Insiders : Override the model used for fast, lightweight utility flows. A fast and inexpensive model is recommended for this setting.

Both settings use Default unless configured, which keeps the GitHub Copilot-provided utility models..

Custom Endpoint provider for BYOK (Insiders)

We now ship a new BYOK provider, the Custom Endpoint provider, that lets you plug any Chat Completions, Responses, or Messages-compatible endpoint into Copilot Chat from a single configuration. It replaces the legacy OpenAI Compatible (customoai) provider, which only supported Chat Completions and is now marked for deprecation.

Screenshot showing the dropdown options for adding a new model in the Language Models editor, with the new Custom Endpoint option.

When you add a model from this provider, you can pick which API family it belongs to (chat-completions, responses, or messages).

Screenshot showing the Custom Endpoint model configuration form with an API type dropdown.

Note: The Custom Endpoint provider is currently in preview and only available in VS Code Insiders.

Integrated Browser

Quickly open HTML files in the Integrated Browser

Previously, previewing an HTML file required installing an extension, which is unnecessary friction for something so common. You can now easily open local HTML files via the Open in Integrated Browser option by right-clicking the file in the File Explorer, or right-clicking the editor tab when the file is already open. You can also select the Preview icon in the editor title bar when an HTML file is active.

Screenshot showing the editor title bar with an HTML file open and an Open in Integrated Browser icon.

Improved experience for adding elements to chat

We have reworked the element selection UI to enable richer functionality and theming support.

Select a range of elements

You can now click and drag to select a range of elements, making it easier to target shared container elements.

Attach elements from context menus

You can now right-click anywhere in a page to quickly attach elements to the chat.

Screenshot showing a context menu opened on an element with an Add Element to Chat item.

Terminal

Agent-aware terminal commands

Command-line tools had no way to tell whether a terminal command was launched by a human or by VS Code's agent flow, which meant progress animations, interactive prompts, and verbose formatting could block or confuse agent sessions.

VS Code now sets a VSCODE_AGENT environment variable for agent-initiated terminal commands. CLIs can check this variable to switch to machine-readable output, suppress progress animations, or skip prompts that would otherwise block the session.

If you maintain scripts or CLIs that already adjust behavior for CI or other agents, you can use the same pattern for commands launched from Copilot Chat.

Running in background indicator for terminal tools

Previously, when a chat terminal command kept running after the tool call returned, the chat UI looked like the command had already finished, making it hard to tell that work was still in progress.

Tool invocations now show Running <command> in background - Show while the terminal is still active. The Show action lets you reveal and focus the underlying terminal. Once the command finishes, the header returns to the normal completed state.

This makes it clearer when a command is still running in the background, especially for async runs or commands that were promoted to background execution after a timeout.

Cleanup of background agent terminals

Previously, when you had a long-running chat session that involved multiple terminal commands, you could accumulate background terminals after each command finished, filling up the terminal list with stale entries and consuming resources.

VS Code now automatically disposes background terminals created by the chat agent when their command completes, while still preserving the command output in the chat UI. If you reveal a background terminal with Show, it stays open so you can continue inspecting or interacting with it.

This keeps terminal lists clean and reduces resource usage over multi-turn sessions.

Broader compression for terminal tool output

Setting: chat.tools.compressOutput.enabled Open in VS Code Open in VS Code Insiders

Commands like pytest, jest, cargo test, tsc, and package installation workflows often produce large volumes of progress output before surfacing the important result, wasting tokens and making it harder for the model to find the relevant information.

Chat terminal tools now compress more kinds of verbose command output before sending it back to the model. The expanded coverage includes common test runners, build tools, linters, Docker commands, and package managers, so repetitive progress information and other low-value output are trimmed more often.

Long terminal runs are now easier for the model to interpret and less likely to spend tokens on boilerplate output.

Sensitive terminal prompts stay in the terminal

Password, passphrase, PIN, or verification-code prompts in terminal commands can pose a risk: the agent could accidentally capture or replay secrets if it tried to handle these prompts itself.

When a chat terminal command reaches a sensitive prompt, VS Code now intercepts it. In default permissions mode, chat shows a confirmation dialog that lets you focus the terminal to enter the secret directly there. In auto-approve flows, VS Code cancels the command and tells the model not to retry or request the secret.

This keeps credentials out of the chat context and prevents the agent from accidentally exposing or replaying sensitive input.

Editor

Quick suggestions default setting change

Copilot's inline suggestions always align with the selection of the suggest control. This is very useful, as you can quickly press Tab twice to accept both the suggestion and the ghost text from Copilot.

However, we've found that as you start to type, in many cases the suggest control pops up and selects the very first (alphabetical) available global symbol that starts with the typed character. This is rarely the text you'd actually type and it also results in Copilot giving you suggestions with that, incorrect, prefix, thus making the experience more noisy.

We've decided to change the default setting for quick suggestions ( editor.quickSuggestions Open in VS Code Open in VS Code Insiders ). If an inline completion provider is available (such as Copilot), then typing letters in the editor no longer automatically triggers the suggest control. In all other cases, the suggest control pops up as before. You can revert to the old behavior by configuring:

"editor.quickSuggestions": {
  "other": "on",
  "comments": "off",
  "strings": "off"
}

Languages

Mermaid diagrams in Markdown preview and Notebooks

We've merged Matt Bierner's Markdown Preview Mermaid Support extension into VS Code as a new built-in extension called Mermaid Markdown Features. This extension adds Mermaid diagram rendering to VS Code's built-in Markdown preview, to Markdown cells in notebooks, and to chats.

Mermaid diagrams can be created using a mermaid fenced code block in your Markdown:

```mermaid
flowchart LR
  Sleep[Sleep] --> Wake{Awake?}
  Wake -->|No| Sleep
  Wake -->|Hungry| Snack[Get treat]
  Wake -->|Not in in Sun?| Move[Move to sun]
  Wake -->|Human is typing| Keyboard[Sleep on keyboard]
  Snack --> Sleep
  Move --> Sleep
  Keyboard --> Sleep
```

Here's what the diagram looks like in the Markdown preview:

Screenshot showing a rendered Mermaid diagram in the Markdown preview.

Rendered Mermaid diagrams also support panning and zooming, which makes larger diagrams easier to inspect without leaving the preview. You can also right-click a diagram to copy its Mermaid source.

YAML frontmatter in Markdown preview

Setting: markdown.preview.frontMatter Open in VS Code Open in VS Code Insiders

We've added options that control how YAML front matter is rendered in the Markdown preview. By default, instead of hiding the preamble, VS Code displays front matter as a table at the top of the preview.

Screenshot showing Markdown frontmatter rendered as a table in the preview.

You can use the markdown.preview.frontMatter Open in VS Code Open in VS Code Insiders setting to choose how front matter appears:

  • table (default): Render front matter as a table.
  • codeBlock: Render front matter as a YAML code block.
  • hide: Hide front matter from the preview.

The rendered frontmatter also has a context menu entry for quickly opening this setting from the preview.

Deprecated features and settings

New deprecations in this release

Upcoming deprecations

Thank you

Contributions to our issue tracking:

Contributions to vscode:

Thank you

Contributions to our issue tracking:

Contributions to vscode:

We really appreciate people trying our new features as soon as they are ready, so check back here often and learn what's new.

If you'd like to read release notes for previous VS Code versions, go to Updates on code.visualstudio.com.