Tokenminning in Grok Bot
Grok Bot runs persistent cloud agents on a shared VM with browser, filesystem, and terminal access. Each turn resends conversation history, Bot description, connector schemas, and computer-use context. Burn shows up as marathon threads, catch-all Bots, broad event routines, and browser clicks where a connector would suffice.
Hub for the Grok Bot cluster. Open a spoke for your query; for the general stack see Where to start.
Quick checklist
- Check Settings → Usage & Billing or Weekly usage — token usage.
- Give each Bot one focused job; avoid a catch-all “General Helper.”
- Scope routines with narrow triggers; test before enabling.
- Prefer connectors over browser computer-use when available.
- Set approval boundaries before granting write access.
Typical impact: 30–50% from scoped first messages and read-only drafts; 20–40% trimming catch-all Bots and stale memory; meaningful cuts from pausing broad event listeners and unused connectors.
How Grok Bot bills
Eligible plans (SuperGrok Plus/Heavy, Cursor Pro+/Ultra, Cursor Teams) include weekly usage. Non-enterprise accounts may see on-demand overage in Usage & Billing. Enterprise teams review usage through the organization administrator or Cursor account page.
Each task can include:
| Component | Notes |
|---|---|
| LLM inference | Model turns, tool planning, summaries |
| Computer use | Browser navigation, terminal commands on the shared VM |
| Connector calls | Structured API access via Plugins |
| Background work | Routines that run while your laptop is closed |
Grok Bot pauses routines after long inactivity if you do not confirm they should keep running — review paused routines when you return.
Guides
| Guide | What you’ll learn |
|---|---|
| Token usage | Weekly meter, usage surfaces, what drives consumption |
| Bots | Focused roles, memory limits, team roster sizing |
| Routines | Skills, schedules, event triggers, test runs, pause controls |
| Computer & apps | Shared VM, connectors vs browser, /workspace files |
| Approvals | Boundaries, Auto Review rules, local-computer execution |
Scope the first message
A strong handoff from Grok Bot get started cuts clarification loops:
- Outcome — what should be finished?
- Sources — which apps, files, or conversations matter?
- Constraints — what must the Bot avoid or ask before doing?
- Deliverable — format and links you expect back
- Review point — when should it stop for you?
Example:
Pull this week’s account list from Salesforce. Skip anyone already in a sequence. Research the top 5 accounts across the web and Slack, draft LinkedIn and email in my voice, and leave drafts to approve by tomorrow morning. Do not send or enroll anyone.
Model routing
When Default Model appears under Settings → General → Agent, default cheap and escalate only on failure — Model routing:
| Tier | Grok Bot choice |
|---|---|
| Read-and-prepare | Default model — research, drafts, summaries |
| Multi-system work | Same Bot with connectors; split only when roles differ |
| Deep debug | Frontier model — last resort on a scoped reproduction task |
Anti-patterns
| Anti-pattern | Why it fails |
|---|---|
| One Bot for every department | Memory and description bloat; hard to reuse skills |
| Broad event listener (“every Slack message”) | Noise, irrelevant runs, wasted computer-use |
| Browser login when a connector exists | Slower, more fragile than structured API access |
| Marathon thread across unrelated tasks | Context compounds; start a new conversation per outcome |
| Routine before a tested skill | Retries and partial failures bill repeatedly |
| Memory as source of truth | Stale facts trigger wrong actions — cite live data |
Related
- IDEs
- Devin — similar cloud-agent metering patterns
- Grok Bot docs