Claude Code Desktop GPU Crash Fix: Stop the Freeze

Posted :

in :

by :

Claude Code Desktop crashes GPU process failure is a confirmed Anthropic regression where opening the app’s in-app Browser/preview pane on sites with bot-detection scripts triggers a fatal Chromium GPU process gone crash (exitCode 101457950) that kills the entire app instantly. For example, verifying a coded webpage against a site running Cloudflare Turnstile can crash Claude Desktop with zero warning dialog.

I’ve spent 33 years in IT, and I still remember the exact moment my Claude Desktop window went black mid-session with no error dialog, no warning, nothing. If you’re reading this because that just happened to you, take a breath: your unsaved session isn’t automatically destroyed. But the path you take in the next ten minutes decides whether you’re back to work in 15 minutes, or stuck for 55 hours like one developer I found in the GitHub bug reports.

Claude Code Desktop crashes GPU process failure is a real, tracked bug — not a sign that your graphics card is dying, and not something you caused by “using the app wrong.” I tested this myself and dug through the official issue threads, and I want to walk you through exactly what’s happening under the hood, why Windows makes it worse than it needs to be, and the precise sequence of steps that actually resolves it.

Claude Code Desktop GPU Crash Fix: Stop the Freeze
Claude Desktop GPU process crash illustration

What Causes the GPU Process Gone Crash? (Quick Answer)

Quick Answer

Since Claude Desktop build 1.24012.1, opening the app’s in-app Browser or preview pane on a site with bot-detection scripts (Cloudflare Turnstile, Akamai, AWS WAF) triggers a fatal Chromium GPU crash — exitCode 101457950 or 0x060C201E — that instantly kills the whole app. It’s reproduced on NVIDIA, AMD, and Intel GPUs across Windows 10 and 11, confirming it’s a software bug, not a hardware fault.

In my own testing, the trigger was almost boring in how consistent it was. I’d ask Claude Code to open its built-in browser to visually check a landing page I’d just coded, the page would load a bot-check script in the background, and within a second or two the entire desktop app would vanish. No crash report, no “Claude has stopped working” dialog — the window just disappears like someone pulled the plug.

That consistency matters. It’s the reason this got escalated from “annoying one-off glitch” to a properly tracked, reproducible bug across three different GPU vendors. GitHub — Anthropic Claude Code Issues

Why “GPU Process Failure” Doesn’t Mean Your Hardware Is Broken

The name of this bug scares people. I’ve seen users assume their GPU driver is corrupted or their card is overheating. It isn’t. Chromium GPU crash issues like this one are almost always a rendering-engine bug, not a physical hardware fault, and this specific case has been confirmed across completely different GPU brands and driver versions. GitHub — Anthropic Claude Code Issues

The real trigger sits inside Chromium’s WebGL handling when it encounters aggressive fingerprinting or bot-detection JavaScript. Here’s the exact chain I traced from the affected users’ logs:

  • The in-app Browser preview pane loads a target site.
  • That site runs a WebGL-based bot-detection or fingerprinting probe.
  • Chromium’s GPU process throws a WebGL CONTEXT_LOST_WEBGL error.
  • The GPU process crashes outright instead of recovering gracefully.
  • The crash takes the entire Electron shell down with it — there’s no fallback.

Why Does the App Get Stuck After Crashing?

This is where the real pain starts. The crash itself is bad enough, but on Windows, the aftermath is often worse than the original problem.

The MSIX Package Gets Flagged “Modified, NeedsRemediation”

After the crash, Windows frequently flags Claude’s MSIX package as MSIX package Modified NeedsRemediation, setting its appxState to 2 and blocking every single relaunch attempt until a manual repair succeeds. GitHub — Anthropic Claude Code Issues I want to be direct about this part because it trips people up badly: clicking “Open” again and again does nothing once the package is in this state.

Worse, on Windows 10 specifically, the built-in Repair function in Settings often only performs what’s essentially a manifest-only re-registration. It doesn’t clear the underlying “Modified” flag, so you can run Repair a dozen times and still be stuck exactly where you started.

A Background Process Locks the Files During Repair

Here’s the detail that cost one user over two full days. Claude Desktop runs a background service called cowork-svc.exe that survives the crash completely intact, even after the main app window disappears. That lingering process holds an active file lock (error code 0x80070020) on core binaries. GitHub — Anthropic Claude Code Issues

If you try to Repair or Uninstall while that process is still running, the operation fails silently or throws a vague error. Most troubleshooting guides skip this step entirely, which is exactly why so many people end up stuck.

Claude Code Desktop crashes GPU process failure trigger diagram
GPU crash trigger chain diagram

How Do You Fix the Claude Code Desktop Crashes GPU Process Failure?

I’m going to give you the sequence in the exact order that worked fastest in the reports I reviewed. Skipping steps, especially step 2 and step 3, is the single biggest reason people waste hours on this.

Step 1 — Confirm the Trigger in Your Log File

Open %APPDATA%\Claude\logs\main.log and check for a GPU process gone entry appearing right after a [PreviewContext] Opened preview user tab line. This confirms the in-app Browser pane specifically caused your crash, rather than something else entirely. GitHub — Anthropic Claude Code Issues

Here is the exact log signature to search for, taken verbatim from the bug reports:

[warn] WebGL: INVALID_ENUM: getInternalformatParameter: invalid internalformat
[warn] The powerPreference option is currently ignored when calling requestAdapter() on Windows.
[warn] WebGL: CONTEXT_LOST_WEBGL: loseContext: context lost
[info] GPU process gone: { type: 'GPU', reason: 'crashed', exitCode: 101457950, serviceName: 'GPU' }

If you see this pattern, or the related 0x060C201E exit code, you’re dealing with exactly this bug and not a separate issue.

Step 2 — Kill All Claude Processes, Including the Hidden One

Open Task Manager and end both claude.exe and cowork-svc.exe. In my experience walking through these reports, skipping that second process is the single most common reason Repair silently fails afterward. GitHub — Anthropic Claude Code Issues It’s easy to miss because it doesn’t show a visible window, so most people never think to look for it.

Step 3 — Run PowerShell Diagnostics Before Repairing

Before you touch the Repair button again, run this in PowerShell:

Get-AppxPackage *Claude* -AllUsers | fl Name,Status

If it returns Modified, NeedsRemediation, stop right there. Don’t bother retrying Repair. One user in the bug thread clicked Repair 25 separate times over 55 hours with zero effect, purely because the package was permanently stuck in that state and only a full uninstall would clear it. GitHub — Anthropic Claude Code Issues Running this one command first is what separates a 15-minute fix from a multi-day outage.

Step 4 — Uninstall and Reinstall Cleanly

If diagnostics confirm the “Modified” state, go to Settings → Apps → Installed apps → Claude → Uninstall. Then download Claude Setup.exe fresh from the official Anthropic downloads page and run a clean install. This forces a proper MSIX re-registration instead of the partial repair Windows attempts on its own.

Step 5 — Prevent Recurrence Without Disabling Hardware Acceleration

A lot of guides recommend disabling hardware acceleration as a first fix. I tested that too, and it’s not a reliable solution here. Disabling it just forces Chromium onto a software WARP renderer, which affected users report still hits the same underlying fault. GitHub — Anthropic Claude Code Issues

What actually works as prevention is simpler: avoid opening the in-app Browser or preview pane on external sites that run bot-detection scripts. Route those specific checks through your regular browser instead, and reserve Claude’s built-in preview for internal or low-risk pages.

Claude Code Desktop crashes GPU process failure fix steps flowchart
Fix sequence for GPU crash recovery

Before You Uninstall, Protect Your Session Data

I want to flag this clearly because it’s the part that generates the most anxiety. Anthropic’s documentation doesn’t clearly state whether local Claude Code session and conversation history survives an uninstall. Claude Code Docs (Official Anthropic) That ambiguity is exactly what caused one affected user to delay their fix for hours out of fear of losing work.

My practical advice: if you have conversation history or session context you care about, export or copy anything critical before you uninstall. Don’t assume it’s preserved, and don’t assume it’s wiped — just don’t gamble on it.

Fix Time Comparison: Repair Loop vs. Diagnostic-First Approach

ApproachTime to ResolutionOutcome
Repeated Repair clicks without checking package statusUp to 55 hours across 25 attemptsFailed repeatedly; package remained stuck in “Modified, NeedsRemediation”
PowerShell diagnostic check, then targeted uninstall/reinstallApproximately 15 minutesResolved cleanly on first attempt

This table alone is the reason I wrote this guide the way I did. The difference between those two rows isn’t luck — it’s one PowerShell command run at the right moment. GitHub — Anthropic Claude Code Issues

A Note on Bad vs. Good Troubleshooting Habits

Bad approach: Repeatedly clicking “Repair” for hours without ever checking the actual package status first. This is exactly what led one user to burn 55 hours across 25 attempts with zero progress, because the underlying package state made Repair fundamentally incapable of fixing the problem. GitHub — Anthropic Claude Code Issues

Good approach: Run the PowerShell diagnostic command first, confirm whether you’re actually dealing with a Modified, NeedsRemediation state, and then choose Repair or full uninstall based on that answer instead of guessing. That single decision point is what collapsed a multi-day outage into roughly 15 minutes for one affected user.

Broader Context: This Fits a Pattern

If you work in software long enough, you start recognizing these patterns. Electron-based desktop apps that embed a browser component are inherently more fragile around GPU handling than native apps, because they’re running a full browser engine inside your productivity tool. When that browser engine hits a hostile script it wasn’t designed to survive, the blast radius is the entire application, not just a tab.

This isn’t unique to Claude Desktop, but the MSIX packaging on Windows makes the recovery process unusually painful compared to, say, a simple app relaunch on macOS. If you want the complete picture on diagnosing and resolving other Claude Code and Claude Desktop issues beyond this specific GPU bug, our complete guide covers the full troubleshooting landscape.

Frequently Asked Questions

Will I lose my Claude Code session history if I uninstall the app?

Local session data preservation through uninstall isn’t clearly documented by Anthropic, so back up any conversation history you need before uninstalling. This exact uncertainty caused one affected user a multi-hour delay out of caution. Claude Code Docs (Official Anthropic)

Is this GPU crash caused by my graphics card or driver?

No. The crash has been reproduced across NVIDIA, AMD, and Intel GPUs with multiple driver versions on both Windows 10 and 11, which confirms it’s an application-level Chromium bug rather than a hardware or driver issue. GitHub — Anthropic Claude Code Issues

Does disabling hardware acceleration fix the crash?

Not reliably. Disabling hardware acceleration forces Chromium onto a software WARP renderer, but affected users still report the same GPU process crash occurring afterward. It’s not a dependable workaround on its own.

What exact error should I look for to confirm this specific bug?

Check your log for GPU process gone with exitCode: 101457950, or on some builds 0x060C201E, often preceded by WebGL: CONTEXT_LOST_WEBGL warnings. This exact signature has been confirmed across at least six independent reports. GitHub — Anthropic Claude Code Issues

How long does the fix actually take once I know the right steps?

Running the PowerShell package-status check before choosing between Repair and Uninstall reportedly cut one user’s recovery time from a 55-hour ordeal down to about 15 minutes. Knowing the right diagnostic step upfront makes all the difference.

Should I report this bug if it happens to me?

Yes. Since this is an actively tracked regression, adding your specific GPU model, Windows version, and log excerpt to the existing issue threads helps Anthropic prioritize a permanent fix rather than a workaround. GitHub — Anthropic Claude Code Issues

References & Sources

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *