[Bug] Failed to start VM: Sandbox not found

[heyalbert]

When trying to sign up for my Claude Code Max subscription I receive the following error: Failed to start VM 8lfjd8: Sandbox not found

I am using macOS 26.4.1

[21st-connector]

🤖 Automation triggered: Implement ISSUE

I'm starting to work on this automatically.

---

View live progress

Powered by 1Code

[21st-connector]

🤖 Automation triggered: Analyze Github Issue

I'm starting to work on this automatically.

---

View live progress

Powered by 1Code

[21st-connector]

Issue Triage

Summary

This error ("Failed to start VM: Sandbox not found") originates from the server-side E2B sandbox infrastructure, not from the desktop app itself. The desktop app connects to cloud-hosted E2B virtual machines for "sandbox mode" chats, and this error indicates the backend failed to provision or locate a sandbox VM during account setup.

Codebase Context

The app has two chat modes (src/renderer/lib/atoms/index.ts:879-887):

• Local mode — worktree-based chats stored in local SQLite
• Sandbox mode — cloud-based E2B sandboxes, accessed via URLs like https://3003-{sandboxId}.e2b.app

Sandbox interactions are handled through backend API endpoints:

• src/renderer/lib/remote-api.ts — remote chat/sandbox API client (getSandboxDiff, getSandboxFile, export endpoints)
• src/main/lib/git/sandbox-import.ts — sandbox state import/export logic
• src/renderer/features/agents/main/active-chat.tsx — sandbox setup status tracking (sandboxSetupStatus, sandboxSetupError)

The error strings "Failed to start VM" and "Sandbox not found" do not appear anywhere in the desktop app source code, confirming this is a server-side error message being surfaced to the user.

Subscription & Auth Flow

The signup flow uses billing methods defined in src/renderer/features/onboarding/billing-method-page.tsx:32-75, with "Claude Pro/Max" (claude-subscription) as the recommended option. User plan data is fetched from the backend via src/main/auth-manager.ts:259-278 (/api/desktop/user/plan).

Related Issues

• Build from source, failed to fork sandbox #31 (closed) — "Build from source, failed to fork sandbox" — different root cause (corrupted node-pty binaries from postinstall script), not related
• [CRITICAL] Sandbox disabled in browser window configuration #103 (closed) — "Sandbox disabled in browser window configuration" — about Electron's renderer sandbox security setting, unrelated to E2B VM sandboxes

No previous issues match this specific server-side VM provisioning error.

Likely Cause

This appears to be a backend/infrastructure issue rather than a desktop app bug. Possible causes:

1. E2B sandbox provisioning failure — the backend couldn't spin up a VM during account setup
2. Transient server-side error — the sandbox was created but became unavailable before the client could connect
3. Account/subscription state mismatch — the signup process may not have fully completed before a sandbox was requested

Suggested Next Steps

1. Retry signup — if this was a transient infrastructure issue, retrying may resolve it
2. Check if the account was created — try signing in again; the subscription may have been activated despite the error
3. Clear app data and retry — on macOS: rm -rf ~/Library/Application\ Support/Agents\ Dev/ (for dev) or ~/Library/Application\ Support/1Code/ (for production), then relaunch
4. Server-side investigation needed — the 1Code team should check E2B sandbox provisioning logs for sandbox ID 8lfjd8 to determine why it wasn't found

This issue likely needs attention from the backend/infrastructure team since the error originates server-side.

[21st-connector]

Issue Analysis

Thanks for reporting this. Here's what I found after reviewing the codebase:

What's happening

The "Failed to start VM: Sandbox not found" error originates from the server-side sandbox provisioning, not from the desktop app itself. During the Claude Code Max signup/authentication flow, the app calls the backend endpoint /api/auth/claude-code/start which creates a cloud sandbox (running on E2B infrastructure) for your session (src/main/lib/trpc/routers/claude-code.ts:174-195).

The error message (Failed to start VM 8lfjd8: Sandbox not found) indicates the server attempted to start a VM with ID 8lfjd8 but the sandbox resource could not be located — this is a server-side/infrastructure issue, not a client-side bug.

Relevant code path

1. User initiates sign-up → startAuth mutation fires (claude-code.ts:178)
2. Desktop app sends POST to {apiUrl}/api/auth/claude-code/start with the desktop token
3. Server creates a sandbox and returns { sandboxId, sandboxUrl, sessionId }
4. If sandbox creation fails, the error propagates back to the desktop UI

Not a duplicate

The two previously closed sandbox-related issues are unrelated:

• [CRITICAL] Sandbox disabled in browser window configuration #103 — Electron's sandbox: false config (security hardening, not VM provisioning)
• Build from source, failed to fork sandbox #31 — Build-from-source postinstall script issue with electron-rebuild

Possible causes

• Transient infrastructure issue: The E2B sandbox backend may have been temporarily unavailable or the specific VM template was not found
• Account provisioning timing: If the Claude Code Max subscription was just initiated, the sandbox quota/allocation may not have been fully provisioned yet
• Region/availability: Sandbox availability may vary by region

Suggested workarounds

1. Retry: Close and reopen the app, then try the sign-up flow again — the sandbox provisioning may have been a transient failure
2. Check subscription status: Verify that the Claude Code Max subscription completed successfully on the billing side before retrying
3. Clear app state (if retries fail):

   # macOS
   rm -rf ~/Library/Application\ Support/Agents\ Dev/
   # Then restart the app

For maintainers

This likely needs investigation on the server/infrastructure side (the /api/auth/claude-code/start endpoint and E2B sandbox provisioning). The desktop app is correctly surfacing the error from the backend — no client-side code change would resolve this.

[21st-connector]

Issue Analysis

What's happening

The error "Failed to start VM: Sandbox not found" originates from the server-side sandbox creation during the Claude Code OAuth flow. When you select "Claude Code Max" as your billing method during onboarding, the app calls the 21st.dev backend (/api/auth/claude-code/start) to spin up a temporary CodeSandbox VM that handles the Anthropic OAuth exchange. The error indicates that the server-side sandbox (VM 8lfjd8) either failed to provision or was not found by the orchestration layer.

Relevant code path

1. Onboarding UI triggers startAuth mutation on mount — src/renderer/features/onboarding/anthropic-onboarding-page.tsx:92
2. tRPC router calls the 21st.dev API to create the sandbox — src/main/lib/trpc/routers/claude-code.ts:174-196
3. If the API returns a non-2xx response, the error is thrown at claude-code.ts:188 and surfaced in the UI via the onError callback at anthropic-onboarding-page.tsx:101-106

The actual sandbox provisioning happens on the 21st.dev server, not in the desktop app. The desktop app simply forwards the request and displays the result.

Root cause

This is a server-side infrastructure issue, not a client-side bug. The "Sandbox not found" error suggests one of:

• The CodeSandbox/E2B VM provisioning service returned an error or was temporarily unavailable
• The sandbox 8lfjd8 was created but terminated or became unreachable before the client could poll it
• A transient infrastructure issue on the 21st.dev backend

Not a duplicate

Previously closed sandbox-related issues (#103 and #31) are unrelated:

• [CRITICAL] Sandbox disabled in browser window configuration #103 was about Electron's renderer sandbox (webPreferences.sandbox: false) — a security config, not VM provisioning
• Build from source, failed to fork sandbox #31 was about electron-rebuild corrupting node-pty binaries during build from source

Suggested workaround

1. Retry — Click the "Connect" button again or restart the onboarding flow. The sandbox provisioning is often transient.
2. Use API key instead — Go back to the billing method selection and choose "API Key" to bypass the sandbox-based OAuth flow entirely. You can use an Anthropic API key from console.anthropic.com.

For maintainers

Some improvements that could help with this class of errors:

• Better error messaging: The raw "Failed to start VM: Sandbox not found" is opaque to users. The startAuth error path at claude-code.ts:188 could surface a more actionable message (e.g., "Authentication service temporarily unavailable, please retry").
• Automatic retry with backoff: The startAuth mutation in the onboarding page fires once on mount. Adding retry logic (e.g., 2-3 attempts with exponential backoff) before showing an error would improve resilience.
• Server-side investigation: Check the 21st.dev backend logs for sandbox 8lfjd8 to determine why it was not found — provisioning failure, timeout, or infrastructure issue.

---

🤖 Generated with Claude Code

[heyalbert]

#Retry: Close and reopen the app, then try the sign-up flow again — the sandbox provisioning may have been a transient failure
Check subscription status: Verify that the Claude Code Max subscription completed successfully on the billing side before retrying
Clear app state (if retries fail):

macOS

rm -rf ~/Library/Application\ Support/Agents\ Dev/

Then restart the app

Tried this. Keeps failing. Active Claude Code subscription. Claude Code with Max subscription works in terminal.
This is a new 1Code user account.

[heyalbert]

Update. New error message when trying to authenticate with Claude: Failed to start auth: Bad Gateway.