Grok Bot vs Grok Build: five names, three programmable surfaces

The map

xAI shipped a great deal in ninety days and named much of it similarly. This table is the whole page in one place; the sections after it add the detail.

Grok Bot: the product with no API

Grok Bot launched on 11 August 2026 in beta. You create named Bots and message them like colleagues. Each Bot works on a persistent cloud Linux VM with a browser, a filesystem and a terminal, and the documentation is specific that the VM is shared: "All of your Bots use the same persistent cloud computer", each gets its own screen on it "without getting separate security boundaries", and you should "treat a login or file placed on the computer as available to all of your Bots".

Bots use connectors and MCP where an API exists, and computer use, meaning driving the graphical interface, for apps and websites that do not have one. Named Bots keep memory, files, browser sessions and preferences across turns. You can walk a Bot through a multi-step path once and it persists that path as a routine to re-run on a schedule or on demand.

What none of that gives you is a way in from your own code. There is no documented public API, no webhook, and no programmatic trigger. Neither the Grok Bot documentation nor any launch coverage describes one. This is a coherent product decision rather than an oversight, because the whole design premise is a human messaging a colleague, but it does mean the phrase "Grok Bot integrations" means connectors that Bots use, not integrations that call Bots.

One platform wrinkle worth flagging because three ranking articles get it wrong in the other direction: the documentation states plainly that Grok Bot is not currently available as a Linux desktop app, while the download page has been reported to serve a Debian package anyway. A Hacker News commenter, gexla, reported exactly that confusion on launch day. If you are on Linux, do not plan around a Linux client existing.

Grok Build: the one you can actually script

Grok Build is xAI's official coding CLI, announced 25 May 2026. Two facts about it are wrong nearly everywhere, and both make it look worse than it is.

• It is not in beta any more. Grok Build reached 1.0 around 7 August 2026 and has shipped steadily since, reaching 1.0.5 by 15 August, and it switched its default model to Grok 4.6 on 13 August. Every article currently ranking for Grok coding comparisons still describes it as an early beta, which was true in May and has not been true for weeks.
• It does not require SuperGrok Heavy. xAI's own announcement puts availability at "SuperGrok and X Premium Plus subscribers", which is the $30 a month tier. Several ranking pages price it at $299 or say there is no individual tier at all. The figure the internet quotes is roughly ten times the real one.

That second point is the single most useful thing on this page for a developer, because it inverts the usual conclusion. Grok Build, the programmable one, is on the cheap tier. Grok Bot, the one with no API, is on the expensive tiers. If you assumed the coding agent was the premium product, you had it backwards.

The CLAUDE.md line is worth pausing on if you are evaluating a switch. A repository already carrying Claude Code project instructions and skills is closer to running under Grok Build than most migration guides suggest, because the CLI reads both instruction formats rather than demanding a rewrite.

The dual authentication path is the thing that trips people up, and it matters for cost. Browser authentication rides a consumer subscription. API-key authentication bills your xAI API account at per-token rates. In CI, or anywhere without a browser, you are on the second path whether you meant to be or not.

The headless surface is the part nothing else on the web documents properly, and it is the reason Grok Build is the answer to "how do I automate this".

One model-level correction while we are here, because it circulated widely in launch-week press and is still repeated: the grok-build-0.1 coding model has a 256,000-token context window, not the 2 million some coverage claimed. If you sized a workflow around the larger figure, resize it. Separately, grok-code-fast-1 was deprecated in May 2026 and retired around 15 August, so any comparison still quoting its benchmark scores is describing a model you can no longer run.

On top of that sits Workflows, announced 23 July 2026: you describe a task in plain language and it fans out across up to 128 parallel agents per run, and up to 1,024 for larger jobs, each with its own focused context. Progress checkpoints let a run pause and resume, /workflows shows live per-agent token tracking, saved workflows become custom slash commands, and a built-in /deep-research verifies claims against sources. Workflows is a Grok Build feature. It is not part of Grok Bot, and conflating the two is one of the more common errors in circulation.

Grok Automations, formerly Tasks

This is the third product people mean when they say they want Grok to do something on a schedule, and most of the pages ranking for it describe a version that no longer exists.

Grok Automations was announced on 16 July 2026, replacing and rebuilding the feature previously called Tasks. It lives at grok.com/automations plus the iOS and Android apps. There are two trigger types: schedules, which run once, daily, on weekdays, weekly, monthly or yearly; and email triggers, which fire on sender, recipient or subject filters. Basic scheduling is free to everyone. Email triggers require SuperGrok at $30 a month.

Automations has no API either. So of the three xAI products that can run work unattended, two of them, Bots and Automations, are driven only through their own interfaces, and one of them, Grok Build, is a CLI you can script. That is the shape of the platform today and it is worth planning around rather than waiting on.

The two remaining confusions

"Grok agent mode" is a query split across two unrelated products, and no page currently resolves it. Imagine Agent Mode arrived in May 2026 at grok.com/imagine and is a creative canvas: image and video work, not code and not automation. Grok Build is the coding agent. If you searched agent mode wanting to ship software, you want Grok Build.

"Grok CLI" is worse, because two entirely different tools ship under the name and neither acknowledges the other. There is grok, the official xAI CLI described above. And there is superagent-ai/grok-cli, a community project with roughly 3,400 stars and active development, offering plan mode, MCP, sub-agents and Telegram remote control, and defaulting to grok-4.3 rather than grok-4.6. Its README contains no reference to the official Grok Build CLI, so anybody arriving from a search for "grok cli" can install it believing it is xAI's own tool.

Neither is wrong to exist. But if a colleague says "use the Grok CLI", ask which one, because the model defaults differ, the authentication differs, and only one of them gets support from xAI. Our Grok CLI guide covers the official tool in detail.

The xAI API, since that is where you end up

Anyone who arrived searching for a Grok Bot API is going to end up here, so it is worth putting the essentials on the page rather than sending you away for them. Everything below is from xAI's own documentation, fetched 22 August 2026.

Two facts about that table deserve emphasis because the pages currently ranking get them wrong. First, you do not need a SuperGrok or Premium+ subscription to use the API, despite at least one well-ranked tutorial implying otherwise. Second, an Anthropic-compatible endpoint is claimed in several secondary sources and we could not verify it: both plausible documentation paths returned 404, so we are not asserting it exists. If your migration plan depends on it, test it before you commit.

Prompt, completion, reasoning and cached tokens all count toward the tokens-per-minute limit, which is a genuine gotcha for anybody assuming cached input is free of quota as well as cheap. Exceeding a limit returns HTTP 429 and the documentation recommends exponential backoff. Non-token charges are billed separately: web and X search at $5 per thousand calls, code execution at $5 per thousand, file attachments at $10 per thousand, image generation at $0.02 each. Batch processing is 20 percent off standard rates on selected models, and priority processing is double.

And the one that catches long-running agents: the 200,000-token threshold applies to the entire request, not incrementally. A 210,000-token prompt bills all of it at $4 input and $12 output rather than the first 200,000 at the standard rate. An agent that accumulates context across a long session crosses that line without any signal that it has.

So what do you actually build against?

If your goal is agents running unattended and controlled from your own code, the honest answer is that you build against a CLI or the raw API, not against a chat product. That is true of xAI and it is true of the category generally.

That last step is where Continuum fits, and it is worth being precise about the relationship. Continuum does not replace Grok Build; it hosts it. It runs Grok alongside Claude Code, Codex, Cursor, Gemini, and OpenCode on a Mac, Linux or Windows machine you own, gives each Code session its own git worktree, gates writes behind a plan you approve, shows live quota gauges, records spend by repository and provider, and is driven from a native iPhone client or the web so the run continues when you close your laptop. The app is free.

If you came here for the Bot side rather than the Build side, what Grok Bot actually is covers the product in depth, and the alternatives guide ranks eight always-on options against it.

Questions people ask

Keep reading

Continue through Grok Bot with Can Grok Bot code? What the shared cloud machine can and cannot do.