Grok Bot

Playbooks

Grok Bot not working — fixes by symptom

Official guidance starts with the least destructive step. Cloud work can continue while the desktop is disconnected. Below groups official troubleshooting, the Cursor getting-started table, and staff posts from awesome “Community & Failure Modes” by symptom — without restating all eighty-nine threads one by one.

Before contacting support, collect: Grok Bot version, OS, exact error, Bot or routine name, time and timezone, request/conversation ID, and what you already tried (Retry / restart / Update). Do not attach passwords, codes, or keys. Email hi@cursor.com.

Can't reach your computer / stuck Reconnecting

Symptom. Chat still works, or you cannot get in at all; messages say Can't reach your computer, Reconnecting, black screen, or blank.

Try first

  1. Fully quit the app (on macOS use menu-bar Quit, not just close the window), then reopen. Official ops guidance starts here for most app issues.
  2. If Retry / Recover computer is offered, take that path first — do not Reset.
  3. Check DNS for *.cursorvm.com. Staff repeatedly point at 1.1.1.1 / 8.8.8.8, and both IPv4 and IPv6 (router RA can keep using ISP DNS). Compare with a phone hotspot on another carrier.

When not to Reset. A temporarily unreachable computer does not mean Bots are gone. If the desktop is black but the same account still works on iOS, do not Reset (that can delete Bots) — wait for an official rebuild. Days of Can't reach after Recover/Reset/reinstall usually mean a stuck hosted box that only staff can restore.

Sources

Reset vs Recover vs Update

Symptom. You want to “rebuild the computer,” Reset stuck on Cleaning up / 50% Starting, or the Reset dialog opens behind Settings and cannot be clicked.

Try first

  1. Update Agent Computer: new image, durable state kept.
  2. Recover Agent Computer: replace the computer when unreachable, still trying to keep durable state.
  3. Reset Agent Computer: return to the latest durable snapshot; unsynced work may be lost. Official docs call this last resort.

When not to Reset. Chat history lives outside the box. Exhausted trial does not delete data — Bots stop answering but Computer view can still export; do not Reset to “save” a trial. Stuck on Cleaning up: backend often already finished — fully quit then Recover, do not Reset again. 50% Starting is a server-side bad state; wait a few minutes; Bot data is usually still there. When a status dialog looks scary, do not click yet — try re-login and phone cellular verification first.

Sources

Weekly usage spills into On-Demand / trial burns instantly

Symptom. Meter at 100%, credits or on-demand charges start, Bot suddenly stops answering, dual limit banners.

Try first

  1. Check Grok Bot weekly usage on the plans screen — not only the Cursor IDE allowance.
  2. If you do not want paid spillover: set On-Demand cap to $0. Note this does not stop promo/referral credits from being drained.
  3. Shrink routine windows; delete idle specialist Bots that poke each other (every Bot↔Bot message counts against weekly usage). Dual banners are usually blocked retries, not double billing.

When not to Reset. Do not assume “stay quiet” in chat stops the meter. Do not treat Pro’s Other Models $20 as the Bot weekly pool. Cloud Agents launched by Grok Bot bill Cursor plan usage separately.

Sources

Plugin OAuth failures (Zoom / Gmail / Notion / Canva / X / GitHub)

Symptom. Connect spins, Invalid redirect, Connected but tools=0, broken Authorization header.

Try first

  1. Follow Connect plugins: Reopen the auth tab; in Teams ask whether Disabled by team admin.
  2. Notion: use Re-authenticate, not Connect again.
  3. Broken GitHub header: long-press account → Remove → sign in again. Canva fails on iOS: connect from desktop; it syncs to phone. When the Gmail plugin is broken, authorize Gmail from Cursor (shared connection).

When not to Reset. Zoom 4700 has no user-side workaround (localhost callback rejected) — do not keep editing Zoom app config. Do not paste bearer tokens into chat. X login risk controls on the cloud computer are a known phenomenon, not proof your account is “broken.”

Sources

Linux / local execution / official vs community packages

Symptom. Linux desktop cannot connect, ListMachines empty, local-exec drops, Always allow still blocks ExternalShell.

Try first

  1. Confirm you installed the official deb/rpm/AppImage from x.ai/bot. Very old unofficial 0.18 builds were rejected; staff once said Grok Bot was not formally on Linux — as of the FAQ fetch day, Linux is listed as an official platform.
  2. Local execution: Settings → Execution on Local Computer. Always Allow lives only on that desktop; iOS can only approve one-shot. Linux 0.30.0 empty ListMachines is a known defect; unlock gnome-keyring/KWallet, kill leftover processes, reopen.
  3. Mac chat works but local shows offline: fully quit. Windows leftover local-exec-daemon: quit from tray, end Grok Bot and daemon in Task Manager, wait about a minute, re-register.

When not to Reset. Do not treat community AUR/COPR/Wine ports as official support. Local stdio MCP is unreachable from the cloud computer. Always allow is not forever-allow.

Sources

White screen / black screen / empty roster (usually not a deleted box)

Symptom. Opens to white, black spinner after install, empty roster after reconnect.

Try first

  1. Fresh install white window: fully quit, rename ~/Library/Application Support/Grok Bot to Grok Bot.bak, relaunch. Reinstall alone does not clear local config.
  2. Empty roster + black/white loading: Agent Computer waking from sleep is slow — nothing was deleted. Do not Reset; wait for the box, fully quit, open again.
  3. Looks like a new user after rebuild: logging in while the computer is still starting shows first-run UI. Fully quit. Do not create a “first Bot” in that window.

When not to Reset. Free-plan black spinner means no hosted computer was provisioned (no seat) — not your network. Team admin role alone does not grant a computer. Stale sessions after password change show “account unavailable.”

Sources

All failure sources (89)

The playbooks above group common symptoms; below keeps all 89 input sources for edge cases not expanded separately. Threads are field evidence, not official promises.