Microsoft vscdeck: A Floating Panel to Watch Every VS Code Copilot Window

Microsoft has put a small but telling tool on GitHub: vscdeck, described as a utility for having overwatch across multiple VS Code instances running GitHub Copilot. It is a Windows panel that sits on top of your screen and tells you which of your coding agents is working, which has finished, and which is waiting for you to answer a question.

The repo is tiny, brand new and has no releases. It is still worth a post, because it names a problem anyone running several coding agents at once already feels: you cannot see all of them.

TL;DR: what vscdeck is

What problem does it solve?

Parallel agents are now normal. Developers open one VS Code window per task, hand each to Copilot, and then lose track. One agent finishes in two minutes and sits idle for an hour. Another stops on a permission prompt and silently blocks. The cost is not compute; it is human attention and wall-clock waiting.

vscdeck turns that into a glanceable strip of buttons. The status colors map to exactly the states you care about: working, done, or needs me. A hotkey brings the right window forward and lands the cursor in the Copilot Chat input, so answering a blocked agent takes one keystroke instead of hunting through the taskbar.

This is the same need that drove tools such as AgentPlane, a local control plane for Claude Code, Codex and Cursor side by side, and the workflow thinking in pstack and Poteto mode. What is different here is that the author is a Microsoft team and the target is Copilot.

How does vscdeck know what Copilot is doing?

This is the interesting engineering detail. According to the README, vscdeck does not hook into VS Code. Instead it reads local Copilot SDK and CLI agent-host session files from the session-state folder under your user profile, incrementally, polling every 750 milliseconds. To avoid reporting stale sessions it checks inuse lock files to see whether the owning process is alive.

For remote work, it can also poll Docker containers and SSH hosts, every three seconds after each poll completes, using the Docker and OpenSSH clients you already have. Nothing is installed in the container or host. Dev containers are matched using the devcontainer.local_folder label, and discovered sources are saved in a RemoteSources setting.

The design is neat because it is non-invasive: a read-only observer of files the agent already writes. It also explains the Copilot SDK angle; our write-up of the GitHub Copilot SDK for multi-platform agents covers the agent runtime that produces those sessions.

How to try it

The README gives no installer, so this is a build-from-source tool:

1. Install the .NET 10 SDK on Windows.
2. Clone the repository.
3. Run dotnet build, or dotnet publish -c Release for a distributable.
4. Launch VSCDeck.exe. The published app needs the .NET 10 Windows Desktop runtime.

The README says no third-party NuGet packages are needed, which keeps the supply-chain surface small. Settings and layout persist in a settings.json under the local app data folder. Optional extras include snapping to screen edges, a notification-area mode and start-with-Windows.

Because it is a brand-new repo, read the code before running it. That advice applies to every utility that reads files containing your prompts and workspace paths.

Limitations the README is honest about

The maintainers list the sharp edges themselves:

• Only agent-host sessions. Traditional extension-host Copilot Chat is not covered, so some setups will show nothing.
• Undocumented files. It relies on local session files, not a stable public Copilot API. A format change can break it without warning.
• Windows focus rules. Windows may deny focus changes, in which case the app flashes the taskbar button rather than forcing focus.
• Unknown beats a wrong guess. Unconfigured remote sources, custom VS Code profiles or data locations, and unreadable sources show Unknown.
• Ambiguous names are reported. Directory-name collisions are flagged, not guessed.

That last pair is a good design principle for agent monitoring generally: a tool that says Unknown is safer than one that confidently shows a green light for an agent that is actually stuck.

What are the privacy and security considerations?

Session files contain what you asked the agent to do and where. A tool that reads them is, by construction, looking at sensitive material. The README warns that the diagnostics file status.json contains local workspace and session descriptions and should be treated as sensitive, so do not attach it to a public issue without scrubbing it.

There is a wider point. As developers run more agents with more permissions, overseeing them becomes a security task as well as a convenience one. Our post on VS Code Remote SSH and agent security shows how agent sandboxes and remote hosts raise questions that a status panel cannot answer. A dashboard tells you an agent is waiting; it does not tell you whether approving its request is wise. If you want a control that blocks risky agent actions rather than just showing state, AgentBeam, the agent security platform from the explainx.ai team, is built for that job.

How does it compare with other approaches?

If you use Claude Code inside VS Code instead, our Claude Code VS Code extension guide covers its own workflow, and the Claude Code vs Cursor vs GitHub Copilot comparison helps decide where your agents should live. vscdeck will not watch those.

A practical workflow for supervising several agents

Whether or not you adopt vscdeck, the habit it encodes is useful. Here is a simple routine for anyone running more than two coding agents at once, based on the states the tool tracks and the official VS Code Copilot documentation:

1. Give each agent its own workspace. One window per task, ideally on its own git branch or worktree, so status maps cleanly to one job.
2. Decide your interrupt policy. Waiting on a permission prompt is the costly state. Decide in advance which actions you will approve quickly and which need a careful read.
3. Batch your check-ins. A glanceable panel lets you check every few minutes instead of staring at one window. Resist the urge to hover.
4. Treat completed as unreviewed. Done means the agent stopped, not that the change is correct. Review diffs and run tests before merging.
5. Keep a log of surprises. If an agent wandered outside its task, note it. Patterns tell you where to tighten permissions.

The deeper lesson is that agent throughput is limited by supervision, not generation. Tools that cut the cost of knowing which agent needs you raise the ceiling on how many you can run, as long as your review discipline keeps pace.

Why a new repo from a top-lab org counts

A repo under the microsoft organization with one star is not a product launch. It is, though, a data point on where Microsoft's Copilot team sees friction: supervising many concurrent agents. Official vendors often prototype such tools as small utilities before folding the idea into the editor itself. If this idea catches on, expect a first-party equivalent inside VS Code rather than a separate WPF app. The README itself does not promise that, and we are not claiming it.

What to watch next

• A first release or installer, which would signal intent to support it.
• Support for extension-host Copilot Chat, which would widen coverage.
• Cross-platform ports, since macOS and Linux developers run just as many agents.
• Any documented, stable session-state format from the Copilot team, which would remove the biggest fragility.

Bottom line

vscdeck is a small, honest, MIT-licensed experiment: a Windows panel that makes parallel Copilot sessions legible. It is not ready for teams to depend on, and it reads sensitive local files, but it shows the next UX layer for agent-heavy development: not a better chat box, but a better way to know which of five chat boxes needs you.

Related reading

• AgentPlane: a local control plane for Claude Code, Codex and Cursor
• pstack and Poteto mode agent workflows
• GitHub Copilot SDK for multi-platform agents
• VS Code Remote SSH and agent security
• Claude Code in VS Code: setup guide
• Claude Code vs Cursor vs GitHub Copilot

Repository details reflect the README on October 14, 2026 and may change.