VS Code 1.147 Is Out: Model Routers, Browser Use and Send Later
VS Code 1.147 lands with model routers, Browser Use, Send Later, and Skills over MCP — and deprecates macOS Ventura, Sonoma, and 32-bit ARM Linux.

Visual Studio Code 1.147 is out now as a stable release, and it leans further into agentic coding: chat requests can be scheduled for later, unsent drafts finally have a home, Dev Container chat sessions can survive a rebuild, and an experimental model router can send different tasks to different models automatically. The update also opens the door to agents driving your actual browser, defaults MCP servers to offering Skills to the model, and — importantly for IT teams — deprecates macOS Ventura and Sonoma support, drops 32-bit ARM Linux, and sets a removal date for the Local agent harness. The release shipped on November 18, 2026, and is available now through VS Code's built-in updater or a fresh download.
What's new in VS Code 1.147, at a glance
Most of 1.147's attention goes to agent sessions and the chat experience around them, continuing the trajectory from 1.146's inbox mode and Copilot Fusion work. The headline additions are scheduling and drafting tools for chat requests, persistent agent sessions inside Dev Containers, an early model-routing system, and a first pass at letting agents control your everyday browser rather than a sandboxed one.
| Feature | Area | Status |
|---|---|---|
| Send Later scheduled chat requests | Agent sessions | Experimental |
| Drafts for unsent chat requests | Agent sessions | Experimental |
| Dev Container chat persistence (Docker volume) | Agent sessions | Experimental |
| Remote environments shared as GitHub Environments | Agent sessions | Available |
| Model routers (YAML task-to-model rules) | Agent experience | Experimental |
| Model comparison panel improvements | Agent experience | Preview |
| HTML previews & native UI responses in chat | Agent experience | Experimental |
| Assisted permissions instructions | Agent experience | Available |
| Screenshot attachments of other app windows | Chat | Experimental |
| Image markup on attachments | Chat | Available |
| Skills over MCP | MCP | Default on |
| Browser Use (drive your own Chrome/Edge) | Browser integration | Preview |
vscode.allowedHarnesses managed setting | Enterprise | Available |
Quick facts
Version: VS Code 1.147 (Stable)
Shipped: November 18, 2026
How to update: Help > Check for Updates, or restart VS Code to pick up the background-downloaded update
Where to read the source: code.visualstudio.com/updates/v1_147
Agent sessions: Send Later, Drafts, and persistent containers
The biggest conceptual shift in 1.147 is treating a chat request like something that doesn't have to be sent the instant you write it. Send Later, gated behind the experimental sessions.sendLater.enabled setting, adds a scheduling option to the send button's menu so a request fires at a specific time instead of immediately. It only works for sessions running "This Computer," requires chat.automations.enabled, and needs the local agent host actually running when the scheduled time arrives — VS Code won't wake a sleeping machine to send it. The agent can also set up a scheduled follow-up from a plain-language instruction, such as asking it to check back on a build in ten minutes.
Drafts, behind sessions.drafts.enabled, is the other half of that idea: a dedicated Drafts section where you can park multiple unsent requests, each remembering its workspace, agent, and model, without anything being sent or started. A draft only disappears once you accept sending it.
For anyone running agent sessions inside Dev Containers, 1.147 adds experimental chat-state persistence across rebuilds. Enabling dev.containers.stateVolume.enabled and chat.agentHost.devContainer.stateVolume.enabled stores session state on a Docker volume rather than inside the container's writable layer, so a rebuild doesn't wipe an in-progress agent conversation. The catch: a container only gets that volume if it was created with the setting already turned on, which means an existing container needs one rebuild before persistence kicks in.
Remote environments also got easier to hand off. A connected SSH host, WSL distribution, or Dev Tunnel can now be shared as a GitHub Environment, so the same remote environment can be reused for agent sessions started from another device or from a supported Copilot app, rather than being tied to the machine that first connected to it. Inbox mode, introduced in 1.146, also picks up an experimental Acknowledged section plus workspace and environment pickers for filtering, behind sessions.list.inboxEnabled.
Agent experience: model routers, comparisons, and assisted permissions
The most consequential new feature for teams juggling multiple models is Model routers, an experimental system behind chat.copilot.modelRouter.enabled. Routers are YAML files stored in ~/.agents/model-routers/ that map task types to specific models, with the routing decision itself handled by a model VS Code calls gpt-6-luna-decisions. The release notes are candid about where this stands: "Model routers are at an early stage, and your feedback will help shape them."
Alongside routers, the model comparison panel introduced in earlier releases moves further out of preview. Comparisons now surface directly under the chat response they came from, and hovering over a past comparison shows the original settings used to generate it — useful for figuring out why two runs of the same prompt produced different answers.
Chat responses themselves get two new experimental rendering modes: chat.htmlPreview.enabled lets an agent reply with an inline HTML preview via /html-preview, and chat.nativeComponents.enabled lets it return native-looking visual components via /vscode-ui. Both are off by default and currently limited to the Agents window.
On the governance side, Assisted permissions adds a chat.permissions.assisted.instructions setting that lets you tell the permission-reviewing model, in plain language, what kinds of actions are safe to approve automatically and what should always stop for a human. Microsoft is explicit that this is additive, not a loophole: those instructions "can't override permission rules, organization policy, or the sandbox." The details panel that shows session and permission information also picks up experimental repositioning options — left, right (the default), above, below, or floating.
Chat: screenshot attachments and image markup
Chat can now reach outside the editor window. An experimental setting, chat.appWindowScreenshots.enabled, adds Add Context > Screenshot App Window so you can attach a screenshot of another running application — a browser tab, a terminal, a design tool — directly into a chat request, using the shortcut Ctrl+Cmd+S on macOS or Ctrl+Shift+Alt+S on Windows and Linux.
Once an image is attached, a new Mark Up option lets you crop it, draw arrows or shapes, and add text before sending — handy for pointing an agent at exactly the part of a screenshot that matters instead of describing it in words. A separate experimental setting, chat.inlineAttachments.enabled, renders attachments inline in the chat input rather than as pills; it ships on by default in Insiders but stays off by default in this Stable release.
MCP: Skills over MCP is now on by default
Following the relevant proposal (SEP-2640), Skills over MCP is enabled by default in 1.147. MCP servers can now offer packaged skills directly to the model, and VS Code asks for your permission the first time a given server tries to use the capability. If you'd rather not have MCP servers surfacing skills at all, the behavior can be turned off with chat.agentHost.mcpSkills.enabled. This builds on the MCP tooling groundwork from recent releases and is likely to matter most for teams running their own internal MCP servers alongside Copilot.
Browser integration: Browser Use lets agents drive your actual browser
The most novel — and most clearly labeled "preview" — feature in 1.147 is Browser Use, toggled with chat.agentHost.browserUse.enabled. Instead of automating a disposable, sandboxed browser instance, it lets an agent work inside your own default Chrome or Edge, using your existing sign-ins, extensions, and session state. It requires three things to be in place: the GitHub Copilot browser extension installed in that browser, a local Copilot harness session running, and Node.js 22 or later on the machine. Because it operates in your real browser, it's worth treating like any other capability that can act with your logged-in identity — scoped and reviewed rather than left on by default.
The same area adds a preview remote HTML preview mode (workbench.browser.enableRemoteProxy) for viewing rendered HTML from a remote environment, along with smaller additions for routing non-localhost links through workbench.externalUriOpeners and dimming the editor behind an open browser overlay via workbench.browser.dimOnOverlay.
Sandboxing and enterprise controls
When a terminal command fails because it hit a sandbox restriction, 1.147 adds a retry flow that re-runs the command with permission scoped to the current session (the default), the workspace, or the user — and the release notes are clear that none of those options permits a command to run outside the sandbox entirely; they only change how long the one-time exception is remembered.
For organizations managing Copilot centrally, the new vscode.allowedHarnesses managed setting can explicitly permit the claude and codex agent harnesses alongside Copilot's own. The list is allow-only: "The list permits the harnesses it names and blocks the others, so an empty list blocks both." Entries must be lowercase, and Microsoft notes that sessions run through those harnesses don't enforce Copilot's own managed controls — a distinction administrators evaluating alternative harnesses will want to read carefully before rolling it out. Managed shell permission rules also gained regex support, so a rule can now be written as, for example, Shell(/^git\s+(push|pull)\b/) instead of listing exact commands.
Deprecations: two macOS versions, 32-bit ARM Linux, and the Local harness's countdown
The part of this release that developers are most likely to overlook is the deprecation list, and it's a longer one than usual. VS Code desktop support for macOS Ventura and macOS Sonoma is deprecated as of 1.147, as is desktop support for 32-bit ARM Linux. Neither is removed in this release, but both are now on notice, which typically precedes a hard cutoff within a release or two. The experimental editor.aiStats.enabled setting for AI usage statistics was also removed outright in this version.
The bigger forward-looking item is one VS Code flagged back in 1.146 and reiterates here: the Local agent harness is deprecated and will be removed in February 2027. Teams that run agent sessions through the local harness rather than a remote environment or an alternative harness have roughly a quarter to plan a transition — worth doing before that date rather than after, given how much of this release (Send Later, Dev Container persistence, Browser Use) assumes a running local or remote agent host rather than a one-off local process.
How to update to VS Code 1.147
VS Code checks for updates automatically and downloads them in the background on Windows and macOS; the update applies the next time you restart the editor, or you can trigger a check manually from Help > Check for Updates (or the Code menu on macOS). Linux users on .deb or .rpm repositories get 1.147 through their regular system package manager, and a fresh installer for every platform is available from the VS Code downloads and updates page. Anyone running the Insiders build will have already seen most of these features land ahead of this Stable release, and the full commit-level changelog is published on the VS Code GitHub release page.
Why this release matters
Individually, none of 1.147's features is a dramatic leap — Send Later and Drafts are small quality-of-life additions, model routers are explicitly an early experiment, and Browser Use is gated behind enough prerequisites that most people won't flip it on this week. Taken together, though, the release reads as Microsoft treating agent sessions as long-running, interruptible processes rather than one-shot chat exchanges: you can schedule them, draft them, persist them across container rebuilds, and hand them off between devices. That's a meaningfully different mental model from "type a prompt, get an answer," and it's the same direction the last several releases — including 1.145's remote agent previews and cron-style automations and 1.144's move to general-availability cloud sessions — have been building toward.
What's next
The clearest action item from 1.147 isn't a new feature at all — it's the clock now running on the Local agent harness, due for removal in February 2027. Anyone with automation, CI steps, or team documentation that assumes a local harness should start testing a remote environment or another supported harness well before that window closes. Beyond that, expect the model router and Browser Use previews to keep evolving quickly given how early-stage Microsoft has described both of them, and expect the usual pattern of experimental settings in this release graduating to on-by-default in 1.148 and beyond, as happened with Skills over MCP this time around. Readers new to the editor can start with Pandromeda's own guide to setting up VS Code and its AI features before diving into agent sessions.
Frequently asked questions
When was VS Code 1.147 released?
VS Code 1.147 shipped as a stable release on November 18, 2026, and is available now through the editor's built-in updater or as a fresh download.
What is Send Later in VS Code 1.147?
Send Later is an experimental feature, enabled with the sessions.sendLater.enabled setting, that schedules a chat request to send at a specific time instead of immediately. It only works for sessions on "This Computer" with chat.automations.enabled turned on, and the local agent host must be running when the scheduled time arrives.
What is Browser Use in VS Code?
Browser Use is a preview feature, behind chat.agentHost.browserUse.enabled, that lets Copilot agents work inside your own default Chrome or Edge browser instead of a sandboxed instance. It requires the GitHub Copilot browser extension, a local Copilot harness session, and Node.js 22 or later.
Is macOS Ventura still supported in VS Code 1.147?
Desktop support for macOS Ventura and macOS Sonoma is deprecated as of VS Code 1.147, meaning it still works for now but is on notice for removal in a future release.
When is VS Code removing the Local agent harness?
Microsoft says the Local agent harness is deprecated and scheduled for removal in February 2027, giving teams that depend on it time to move to a remote environment or another supported harness.
What does Skills over MCP do in VS Code 1.147?
Skills over MCP lets MCP servers offer packaged skills directly to the model inside VS Code. It's enabled by default starting in 1.147, though it can be turned off with chat.agentHost.mcpSkills.enabled.
Sources
- VS Code 1.147 Release Notescode.visualstudio.com
- VS Code Updates & Downloadscode.visualstudio.com
- VS Code Dev Containers Documentationcode.visualstudio.com
- VS Code MCP Servers Documentationcode.visualstudio.com
- VS Code 1.146 Release Notescode.visualstudio.com
- VS Code 1.147.0 on GitHubgithub.com
Felix Moreau writes Pandromeda's software coverage and how-to guides. He covers Windows, macOS and Linux updates, the apps people rely on, emulators and developer tools, and turns official documentation into clear, numbered steps that work on the current version.


