Fix: Probleme mit dem Antigravity v1.21.6 Update

The Antigravity v1.21.6 update rolled out in late March 2026 with sweeping changes to sandboxing, MCP authentication, and tool execution. While the changelog advertised security improvements, the release introduced a wave of regressions that broke core workflows for a significant number of users. This guide documents every confirmed issue, its root cause, and the fix — ordered by severity.

What Changed in v1.21.6

Before diving into fixes, it helps to understand what v1.21.6 actually shipped. The release was primarily a security and compliance update, but several changes had unintended side effects:

• Linux sandboxing (new): A new mandatory sandbox mode for Linux and WSL2 environments. This is the single biggest source of breakage — the sandbox restricts file system access and subprocess execution by default, with no opt-out toggle in the initial release.
• MCP authentication overhaul: The MCP server auth flow was rewritten to use a new token-exchange mechanism. Existing auth tokens from v1.20.x are now invalid, and the new auth page has its own set of bugs.
• Tool execution engine refactor: The internal engine that handles write_to_file, execute_command, and other built-in tools was refactored for the new sandbox. This broke tool calls on multiple platforms.
• Feature deprecations: Follow-along mode and the playground environment were removed entirely (not just hidden — removed from the codebase).
• Auto-updater changes: The macOS auto-updater now runs more aggressively, making it harder to stay on an older version.

The combination of mandatory sandboxing, auth changes, and tool engine refactoring created a situation where users on Linux, WSL2, and Windows hit multiple issues simultaneously. The sections below address each one individually.

Fix: Sandbox Mode Locking You Out (WSL2 / Linux)

This is the most reported issue from v1.21.6. After the update, WSL2 and native Linux users find themselves locked into a restricted sandbox where the agent cannot write files, execute terminal commands, or access project directories outside a narrow allowed path. The error typically appears as:

Root cause: The v1.21.6 sandbox defaults to a minimal allowlist that only includes the ~/.antigravity directory. Your actual project directories are not added automatically on Linux/WSL2, unlike macOS where the app bundle handles path registration.

The Fix

You need to explicitly configure the sandbox allowlist. Open (or create) the file ~/.antigravity/sandbox.json:

Replace YOUR_USERNAME with your actual Linux username. After saving, restart Antigravity completely:

Wenn nach der Konfiguration der allowlist weiterhin Berechtigungsfehler auftreten, hängt Ihr Problem möglicherweise mit der allgemeinen "version no longer supported"-Sperre zusammen — siehe unseren Leitfaden zur Behebung von & for that scenario.

Fix: write_to_file Tool Broken

After updating to v1.21.6, the write_to_file tool silently fails — the agent reports success, but the file is never actually written or modified. This affects both the built-in file writing tool and any MCP server that relies on file system access.

Root cause: The tool execution engine refactor introduced a race condition where the sandbox permission check completes after the write operation is attempted. The write is blocked, but the success callback fires anyway because it was registered before the permission check.

Workaround

Until a patch is released, the most reliable workaround is to use the execute_command tool with a shell redirect instead of write_to_file. You can instruct the agent to do this by adding a rule to your project's .antigravity/rules file:

This is not ideal for large files, but it works consistently for most development workflows. The agent will use cat or echo with output redirection, which goes through the command execution path rather than the broken file-write path.

Fix: Stuck on "Running Command" Indefinitely (Windows)

Windows users report that after v1.21.6, the agent gets stuck displaying "Running command..." with no output and no way to cancel. The spinner runs forever, and even starting a new conversation does not clear the state.

Root cause: The new sandbox layer intercepts subprocess calls on Windows but does not properly handle the case where the default shell (cmd.exe) is invoked without an explicit path. The sandbox waits for a permission grant that never arrives because the request is malformed internally.

The Fix

Set the default shell explicitly in your Antigravity settings. Open Settings > Terminal > Default Shell and set it to the full path:

Oder, falls du PowerShell bevorzugst:

Der entscheidende Punkt ist die Verwendung des absoluten Pfads anstatt nur cmd oder powershell. Beende Antigravity nach dieser Einstellung vollständig (prüfe im Task-Manager, ob keine Hintergrundprozesse mehr aktiv sind) und starte es neu. Falls der Agent mitten im Gespräch stecken geblieben ist, musst du eventuell eine neue Konversation beginnen — die alte könnte in einem fehlerhaften Zustand verbleiben.

Bei allgemeineren Problemen Antigravity not responding fix guide.

Fix: PowerShell Execution Errors

A related but distinct issue: even when commands do execute on Windows, PowerShell-specific syntax fails with errors like The term 'X' is not recognized as the name of a cmdlet or execution policy blocks. This happens because v1.21.6 changed how the agent spawns shell subprocesses, and PowerShell's execution policy is not inherited correctly in the new sandbox.

The Fix

Add the following to your Antigravity terminal configuration (Settings > Terminal > Shell Args):

This tells PowerShell to skip the execution policy check and profile loading, both of which cause problems in the sandboxed environment. If you are a cmd.exe user rather than PowerShell, switching to cmd as your default shell (see previous section) avoids this entirely.

Fix: Tool Call Approval UI Disappearing

In v1.21.6, when the agent requests permission to run a tool (file write, terminal command, etc.), the approval dialog briefly flashes on screen and then vanishes. The agent then sits idle, waiting for approval that the user cannot grant because the UI element is gone. There is no way to approve or deny the request.

Root cause: The approval UI component's z-index conflicts with a new overlay layer added in v1.21.6 for sandbox status notifications. The approval dialog renders behind the overlay and is immediately obscured.

Community Workaround (via Global Rules)

Until the UI bug is patched, you can bypass the approval dialog entirely by configuring the agent to auto-approve tool calls. Add this to your global rules file:

An alternative is to run Antigravity with the --auto-approve flag from the command line, which scopes auto-approval to a single session rather than making it a permanent global rule:

Fix: MCP Auth Page Broken

The v1.21.6 MCP authentication rewrite broke the OAuth flow for third-party MCP servers. When you click "Authenticate" on an MCP server that requires login (like GitHub, Supabase, or Vercel), the auth page either loads a blank white screen or throws a redirect_uri_mismatch error.

Ursache: Der neue token-exchange-Endpunkt erwartet ein anderes Format für die Callback-URL, aber die MCP-Serverdefinitionen verweisen noch auf das alte Format. Dies ist eine serverseitige Konfigurationsdiskrepanz, die vom Antigravity-Team behoben werden muss.

Workaround

Sie können MCP-Server manuell authentifizieren, indem Sie Token außerhalb von Antigravity generieren und diese in Ihre Konfiguration einfügen:

1. Erstellen Sie ein Personal Access Token (PAT) im Dashboard des MCP-Anbieters (z. B. GitHub Settings > Developer settings > Personal access tokens).
2. Fügen Sie das Token direkt zu Ihrer mcp_config.json using environment variable substitution:

3. Set the environment variable in your shell profile (~/.zshrc, ~/.bashrc, or Windows environment variables).
4. Restart Antigravity to pick up the new config.

For a complete walkthrough of MCP server configuration and common errors, see our MCP initialize EOF fix guide.

Fix: Terminal Commands Not Working on Linux Mint

Linux Mint users specifically report that all terminal commands fail in v1.21.6 with a generic "command execution failed" error, even basic ones like ls or pwd. This does not affect Ubuntu, Fedora, or Arch — only Linux Mint and its derivatives.

Root cause: Linux Mint ships /bin/sh as a symlink to dash, and the v1.21.6 sandbox shell detection incorrectly identifies dash as an unsupported shell. The sandbox then refuses to spawn any subprocess.

The Fix

Override the shell in your Antigravity settings. Open Settings > Terminal > Default Shell and set:

If you prefer zsh, use /usr/bin/zsh instead. The key is explicitly pointing to a supported shell binary rather than relying on /bin/sh.

Alternatively, create a symlink override on your system (this fixes it system-wide for all applications that have the same assumption):

So führen Sie ein Downgrade auf v1.19.6 / v1.20.5 durch

Falls die oben genannten Fixes Ihre Probleme nicht lösen oder Sie auf entfernte Funktionen angewiesen sind (siehe nächster Abschnitt), ist ein Downgrade die zuverlässigste Lösung. Die zwei empfohlenen stabilen Versionen sind v1.20.5 (die aktuellste Version vor der Sandbox-Überarbeitung) und v1.19.6 (die letzte Version mit Follow-along-Modus und Playground).

Schritt 1: v1.21.

macOS

Windows

Linux

Step 2: Install Your Target Version

Download the specific version installer from your Google account's Antigravity release archive. Log into your account dashboard, navigate to the Downloads section, and select the version you want.

Step 3: Bypass the Auto-Updater (Critical)

This is the step most people miss. After installing an older version, Antigravity's auto-updater will immediately detect that a newer version exists and force-update you back to v1.21.6 — sometimes within minutes of launching.

macOS (block the updater):

Windows:

Linux:

If you encounter a "version no longer supported" message after downgrading, that is a separate server-side enforcement — see our version no longer supported fix guide für Workarounds.

Die entfernten Features: Follow-Along Mode & Playground

Zwei Features, auf die sich viele Nutzer verlassen haben, wurden in v1.21.6 ohne Deprecation-Warnung stillschweigend entfernt:

• Follow-Along Mode: Previously let you watch the agent's cursor movement and file edits in real time. This was removed for "performance optimization" according to the internal changelog. The agent still works in the background, but you no longer see a live view of its actions.
• Playground: Die Sandbox-Playground-Umgebung, in der man Agent-Aktionen testen konnte, ohne das eigentliche Projekt zu beeinflussen. Ironischerweise wurde sie in demselben Update entfernt, das obligatorisches Sandboxing einführte. Die Begründung war, dass der neue Sandbox-Modus den Playground überflüssig macht, obwohl viele Nutzer anderer Meinung sind.

Aktuell gibt es keinen Workaround, um diese Features in v1.21.6 wiederherzustellen. Falls sie für Ihren Workflow entscheidend sind, ist ein Downgrade auf v1.19.6 (die letzte Version, die beides enthält) die einzige Option. Beachten Sie, dass v1.20.x bereits damit begonnen hatte, den Follow-Along Mode hinter einem Feature-Flag als veraltet zu markieren, daher ist v1.19.6 genau die Version, die Sie benötigen.

Falls Ihre Antigravity-Instanz abstürzt, anstatt nur Features zu vermissen, könnte die Ursache eine andere sein — prüfen Sie unseren Guide zur Behebung von Server-Abstürzen für spezifische Fehlerbehebung bei Abstürzen.