Grok Bot (Ultra) unreliable for business use — credit balance load failure, Slack drops, router bot degradation under load, insufficient native memory/project/task tracking

Support Bug Reports
,
1

Where does the bug appear (feature/product)?

Grok Bot

Describe the Bug

Grok Bot on Ultra plan is unreliable for business use. Credit balance fails to load (dashboard and Billing pages show “Couldn’t load the credit balance”) even though weekly included usage correctly shows only 10% used and resets the next day. Bot then incorrectly blocks or prompts as if the prepaid balance is exhausted.

Slack integrations drop connections, delay replies, and cause loops. Central router/CEO-style bots degrade under multi-project and higher routine load. Native memory is insufficient for durable project/task tracking, requiring external file-based state management and extra maintenance processes.

Steps to Reproduce

Be on Ultra plan with remaining weekly included usage (e.g. 10% used, reset upcoming).
Open dashboard or Billing & invoices page — credit balance fails to load with “Couldn’t load the credit balance” error.
Use Grok Bot — it surfaces “prepaid balance is used up” / Buy Credits despite available weekly pool.
Hard refresh, Try Again, and logout/login do not clear the credit balance error. Only a full Grok Bot app restart does.
For Slack: trigger event-driven routines or staff questions via Slack; observe connection drops, delayed responses, or loops.
For router degradation: run a central bot with multiple routines and active multi-project state; observe reduced reliability and performance until load is manually reduced and state is externalized.

Expected Behavior

Credit balance should load reliably when the Ultra weekly included usage is available.
Bot should draw from the remaining weekly pool before requiring or blocking on prepaid credits.
Slack connections should remain stable without drops or loops.
Router-style bots should handle realistic multi-project and routine loads without degradation.
Native memory and context should support durable project and task tracking without requiring external file systems and dedicated maintenance bots.

Operating System

MacOS

Version Information

Version: 0.72.1
Built: 2026-10-20T19:18:02.695Z
Release Track: stable
OS: darwin

For AI issues: which model did you use?

N/A (or Grok Bot default routing)

For AI issues: add Request ID with privacy disabled

N/A (or provide if available from affected sessions)

Additional Information

Ultra is the highest non-Enterprise tier.
Weekly included usage correctly displayed at 10% used (resets Oct 22) while credit balance endpoint failed to load.
Slack drops and loops affect staff question handling.
Central router bot degraded under ~14 routines + multi-project state; reliability improved only after reducing to ~5 routines and offloading work.
Native memory proved insufficient, so durable state had to be externalized to shared-computer files (project indexes, rules, logs, handoffs) with standing instructions for all bots to read/write them, plus a dedicated maintenance bot to keep the external state consistent.
These are reliability and capability gaps that make the product unsuitable for ongoing business use in its current state. Happy to provide more logs or high-level examples.

Does this stop you from using Cursor

No - Cursor works, but with this issue

3

+1 with this one, I cant use it at all since 2 days ago

5

Note: The credit-balance load failure (“Couldn’t load the credit balance”) on the dashboard and Billing pages aligns with the Oct 21 service degradation affecting billing dashboards (status notice ~15:46 UTC).

8

Hey @Joe_LK,

Thanks for the detailed report, and yes, you spotted it in your note. The “Couldn’t load the credit balance” error is due to a temporary issue on our side while reading billing data. It isn’t your account and nothing on it needs to change. The team is working on it, and once it’s resolved, the dashboard and Billing page should display your balance again. Please don’t buy credits just to get past that screen.

The “prepaid balance is used up” prompt while your weekly usage shows 10% looks like a separate problem we’re already tracking. I’ve shared your report with the team.

For the Slack problems, could you share a couple of examples: roughly when they happened (with your timezone), which bot or routine was involved, and whether the bot missed a message, replied late, or replied more than once? For the router bot, it helps to know how many bots and routines were active and what “degraded” looked like (slower replies, skipped routines, wrong routing, or errors).

Ddurable memory and project and task tracking across bots is on the team’s radar, and your setup (shared files plus a maintenance bot) is a good example of why it matters. For now, that kind of file-based state is still the most reliable way to keep long-running projects consistent.

@Jonathan_Pilatti, what you’re seeing may be a different problem. Could you open a separate thread with what happens when you try to use it and the error you see? That way it can be looked at on its own.

10

Thanks so much for the quick and clear update (as usual), Collin — really appreciate it

Slack Bot: Here are some concrete examples. All times are US Central (CT).

How we use it: our staff aren’t very technical, and Slack is how they reach our Grok Bot setup. Through our Slack bot “Pepe Silvia” they ask questions, and they also kick off workflows we’ve built on the back end. Those include writing and refreshing article drafts, building briefs from our research dashboard, approving or cancelling jobs, price rechecks, compliance checks and reports. Behind it is a shared “brain” of project files, rules and tracker data, plus about 14 agents that split the work. Staff never open Grok Bot. They just @mention Pepe in a work channel or reply in its thread, and the right agent does the job and posts back.

Setup: Pepe Silvia is one Slack app with a custom bot token, connected to a personal Grok Bot. Because a mention of a custom bot doesn’t wake a personal Grok Bot, we built our own wake path. A Socket Mode listener POSTs to a webhook routine, and a backup poll routine catches anything missed.

Examples:

  1. Stuck “working” status, no reply (late or missed). Aug 17, around 11:33 PM CT, in a work channel. A teammate asked the bot a question. The listener set the “is working on it…” status, but the agent run that should answer had already exited, so no reply went out until I pinged it the next morning.
  2. Status loop, no reply. Aug 21, the 1-minute poll routine. Polls stacked on top of a run that was still in progress, so the status kept flashing and no reply was posted. We added an overlap lock to stop it.
  3. Listener gaps (missed or late). Aug 22 and Sep 8, overnight. The listener and wake jobs on the agent’s computer stopped, and mentions went unanswered for hours. That’s why we ran a 5-minute health-check routine, and on Oct 20 we moved the listener to our own always-on server.
  4. Replied more than once. Oct 20, 12:51 PM CT. I @mentioned the bot in a work channel and got two replies. Two of our agents both had listeners on the same channel. Earlier, a “whats up?” message got the same double reply. We fixed it by making one agent the only listener.

Router bot: our main bot routes work to about 13 other agents, which run roughly 20 or more routines between them, several of them hourly. On Oct 20 several routine runs failed with “no Grok Bot usage available”, so they were simply skipped. In general, “degraded” has meant slower replies as chats got long, two agents acting on the same Slack message, and a status reported from memory instead of being checked live.

What would fix this for us: Team Bots look close, but they only answer people who are on our Cursor team and have linked their Slack accounts. They also hand work only to that person’s own bots. Our staff don’t need their own Cursor seats or bots, since they’d never open the app. They need Slack access to the setup I’ve already built: my agents, routines, connected apps and project files.

Is there, or could there be, a way for a bot to answer anyone in our Slack workspace and run that work as me, on my account and usage? That includes handing tasks to my other agents from a channel thread. A per-bot setting like “unlinked workspace members act through the owner” would let us drop our custom listener entirely and stop the misses and double replies. Other suggestions?

Happy to provide more info if there’s anything specific you’d like to see.

Glad to hear durable memory / cross-bot project tracking is on the radar. Happy to share more about our current file-based setup if it’s useful.

Thanks again for you help, overall Grokbot has been great.

14

Hey @Joe_LK,

Short answer on the Slack question: not today. A personal Bot connected to Slack only answers its owner. A Team Bot answers people on your Cursor team who have linked their Slack account, and each reply uses the Grok Bot usage of the person it’s answering, not the owner’s. Anyone who hasn’t linked an account gets a prompt to link one, and there’s no setting that lets them work through the owner’s account, Bots and usage.

Letting people without a Cursor seat talk to a shared Bot in Slack is a request the team is already tracking! For now, a custom listener like yours is the way to do it. The changes you made (one listener per channel, an overlap lock on the poll routine, and moving the listener to an always-on server) address the misses and double replies you described.

On the Oct 20 skips: the routine runs that stopped late that morning (around 11:00 to 11:30 AM CT) line up with the weekly included usage being fully used at that point, with no credits to fall back on. Routines stop like that when usage runs out, and it’s separate from the “prepaid balance is used up” prompts you saw today.

Today’s prompts look like they came during the billing problem I mentioned earlier, which the team has now mitigated. If Grok Bot still asks for credits after you restart the app, reply here and we’ll take another look.

15

Thanks for the clear breakdown, Collin

Good to know the request for letting people without a Cursor seat talk to a shared Bot in Slack is already being tracked. I think that would meaningfully increase business usage, since a lot of teams have staff or collaborators who don’t need full Cursor seats and just aren’t that technical, but still need to interact with the Bot through Slack.

Also thanks for clarifying the Oct 20 routine skips (weekly usage exhausted with no credits to fall back on) and confirming that today’s credit prompts were tied to the billing dashboard issue that’s now mitigated. That lines up with what we saw.

I’ll keep an eye on it after the next app restart and will reply here if the credit prompts come back.