News

VS Code v1.110 Insiders: AI Agents Gain Native Browser Access and Global Instructions

• By David Ramel
• 02/25/2026

The headline AI feature in the February 2026 Insiders notes for Visual Studio Code 1.110 is "Native browser integration in chat," described as enabling you to interact with page elements, capture screenshots for visual debugging, and access real-time browser console logs directly by AI agents.

This and other new features and functionality will likely hit VS Code 1.110 stable next week.

The same release notes reference a related feature request that provides more detail on what this is intended to replace. In that issue, the author says Copilot's browser support "relies on external Model Context Protocol (MCP) tools for browser operations," and argues that depending on external tools can produce "inconsistent implementation," reduced efficiency from switching tools or managing dependencies, and a "fragmented user experience" when web development tasks require bouncing between the editor, browser tooling, and AI.

In the proposal, "native" integration is framed as bringing the browser-side work into the Copilot experience: direct interaction with page elements (HTML/CSS/JavaScript), full or partial screenshots for visual debugging, and real-time access to console logs, including signals like JavaScript errors and network requests. The issue is closed and labeled "insiders-released," aligning with the capability now appearing in the Insiders notes.

I used my Insiders build to open up my Visual Studio Magazine site from Chat (above) and asked Copilot if this illustrated the new capability. It responded:

Yes, this illustrates it. When I opened visualstudiomagazine.com using the open_browser_page function, it displayed the page in VS Code's integrated browser--which is an example of the native browser integration in VS Code. This allows web content to be viewed and interacted with directly within the editor environment, which is part of the enhanced chat and agent capabilities in recent VS Code versions.

This integration makes it possible for agents and chat assistants to browse web pages, interact with content, and perform web-based tasks without leaving the editor, streamlining workflows for tasks like research, documentation lookup, and content inspection.

More Agent and Chat Workflow Changes
Beyond the browser work, the Insiders notes stack up a set of changes that focus on how developers steer agents and manage conversations while work is in progress.

• Background agents: Support for slash commands for background agents.
• Branching conversations: The /fork command to create a new session inheriting the current conversation context.
• Prompt queueing: The ability to queue a second prompt while the first is still running, then have it processed automatically when the current task completes.
• Subagents and questions: The askQuestions tool working in subagent contexts, and visible rendering of subagent invocations during streaming for Claude Agent sessions.

There are also smaller chat interaction changes aimed at reducing friction, including letting inline chat send "Add selection to chat" into the active chat queue, and tweaks to chat tips behavior (including not showing tips when the user is not signed in, and more control over dismissing or hiding tips). The release notes also include a redesigned model picker with sections for recent models and search functionality.

Models, Context, and Instruction Handling
Several AI-facing updates in the notes focus on model selection, context management, and how instructions are discovered and applied.

• Instruction discovery: The ~/.copilot/instructions directory automatically added to configured instruction load paths. This gives Copilot Chat a user-level place to pick up instruction files without having to add that directory to your settings. The intent is to make "global" instructions easier to reuse across projects, alongside any repo-scoped instruction files.
• Context window visibility: Context window usage controls shown for Claude Agent sessions, and a setting to hide the indicator in the chat input area. This exposes how much of the model's context window is being consumed during a session. If you find the indicator distracting, the new setting lets you remove it from the input UI.
• Actionable context controls: A context usage control that can recommend ways to manage context and allows manually triggering conversation compaction. When the session approaches the context limit, the control can surface guidance for keeping the conversation within limits. It also adds a manual “compaction” action to summarize older conversation content to free up context space.

One tooling change is also framed in AI-utility terms: when a tool returns a large amount of data, output can be written to a temporary file on disk instead of being truncated, which the release notes describe as improving response accuracy and reducing token usage.

MCP and Sandboxing
For developers and organizations watching extensibility and controls around agent tooling, the release notes include two MCP-related items.

• Claude Agent support for MCP (Model Context Protocol) servers, with MCP servers installed via VS Code or the Claude CLI "automatically picked up."
• A sandbox option for local MCP servers using stdio transport, described as adding an isolation boundary for file system and network access.

When the Goodies Hit Stable
The 1.110 items above are currently tracked in the February 2026 Insiders notes and are expected to consolidate into the next stable release. In the VS Code team's February 2026 iteration plan, the timeline is described as a four-week iteration with endgame running Feb. 23-27, 2026, and the plan states, "We will ship in early March."

Non-AI Roundup
Outside the AI changes, 1.110's Insiders notes include a batch of platform and usability updates.

• Terminal: Support for the Kitty graphics protocol (inline images), pixel-dimension reporting during terminal resize, and additional terminal control/reporting features.
• External terminal: Support for Ghostty as an external terminal on macOS and Linux.
• Windows IT/security: Win32 version metadata added to shipped executables to enable AppLocker rules that require signed executables with product metadata.
• Update and UX: Progress indicators for downloads/installs, metered network awareness for postponing updates, and multiple accessibility improvements across dialogs and chat.

As with any Insiders cycle, individual items can shift between the initial stable release and follow-on recovery updates, but the throughline of this iteration is consistent: more agent capability, more agent control surfaces, and tighter integration between AI chat and the tools developers already use for day-to-day debugging.