By ByteBot
Grok Bot is not a coding assistant. It’s a persistent digital worker that signs into your apps, learns workflows by watching you once, and then executes them 24/7 — on its own cloud computer, with memory that compounds across sessions. On September 3, xAI opened enterprise-wide deployment with audit controls, SCIM provisioning, and network policies. The governance gap is already visible: action recording ships off by default, there’s no per-org spending cap, and the security boundary is the account — not the individual Bot.
This Is Not Grok Build
If you’ve been tracking xAI’s agent stack, the naming is intentional: Grok Build is the coding harness — the thing that reads your repo, writes code, and opens PRs. Grok Bot is different. Each Bot gets its own persistent Linux VM, browser sessions, file storage, and memory that compounds rather than resets. It runs on Grok 4.6’s 500K context window, can observe a workflow once and re-execute it automatically, and multiple Bots can message each other to coordinate across systems.
The closest analogy: Cursor Cloud Agent is an isolated VM that clones your repo and opens a PR. Grok Bot is a persistent office computer that stays signed into your apps after you close the laptop. That’s a different category entirely.
What the Enterprise Tier Adds
The September 3 enterprise launch unlocks a set of controls not available on self-serve Teams plans:
- Organization-wide enable/disable switch — disabling blocks access without deleting member computers
- Network Controls — domain and IP allowlists for all bot traffic; Teams defaults to allow-all with no configuration options
- SCIM 2.0 — automated provisioning and deprovisioning through your identity provider
- Action Recording — captures what bots actually do (disabled by default)
- OpenTelemetry Export — stream sanitized bot events to your own SIEM or collector
- MCP Allowlist — restrict which MCP server endpoints bots can reach
- Audit Logs — admin, security, and auth events, plus Grok Bot control-plane events
- Computer Management — locate, review, and terminate any member’s bot instance
Existing Grok or Cursor Enterprise customers get a two-week free org-wide trial and can invite every employee during that window.
The Governance Gap You Need to Read
Here’s what the official documentation quietly acknowledges: most of the controls that matter are opt-in, not default.
Action Recording is off by default. Cloud Agent delegation is on by default. Auto Review is active but individual members can disable it. There is no spending cap for unattended agents. Admins cannot restrict which model serves requests (described in the docs as “not planned”). Team-level local execution limits are “coming soon.”
The shared infrastructure risk is the one that stands out most. xAI’s own security docs state it directly: “Do not use separate Bots as a security boundary.” Every Bot on an account shares one VM — files, browser sessions, and installed connectors are account-wide with no per-Bot isolation. A credential stored by one Bot is accessible to all of them. And the egress IP is shared across customers, so you can’t distinguish your org’s traffic from other tenants on the same ranges.
There’s also a mechanics issue with approvals: they stop future actions, not past ones. A Bot can execute a sequence, then pause for human approval. The damage, if any, has already happened before the review queue.
Access and What It Costs
Grok Bot access scales with your Cursor or SuperGrok plan. The notable price drop: it launched at $200/month (SuperGrok Heavy only) on August 11, fell to $60 on August 21, and landed in Cursor Pro at $20/month on August 26. Enterprise controls require an Enterprise plan — they don’t trickle down to Teams Standard or Premium.
- Cursor Pro: $20/month (includes Grok Bot)
- Cursor Teams Standard: $40/seat
- Cursor Teams Premium: $120/seat (SSO included)
- SuperGrok: $30/month | SuperGrok Heavy: $300/month
What to Do Before You Deploy
If you’re evaluating Grok Bot for enterprise use, the configuration defaults are the first thing to address:
- Enable Action Recording immediately — it captures nothing until you turn it on
- Set Network Controls to a domain allowlist rather than the default allow-all
- Set local execution to “Never allowed” unless your use case specifically requires it
- Write narrow Auto-Review rules — broad browser-access approvals create uncontrolled surface area
- Treat the entire account as the security boundary, not individual Bots
- Check the 200K context pricing cliff on Grok 4.6 — crossing it roughly doubles the cost of the entire request
The category is real. Autonomous enterprise agents that sign into your tools and handle multi-step workflows are no longer a proof-of-concept. The question now is whether your IT policy is ready for workers who never clock out and can’t be individually audited without opting into the feature that captures their actions.
