Grok Bot is an agent that uses your apps by hand.

A note on sourcing

This post is more heavily caveated than most of ours, because Grok Bot is badly documented for a product this visible. The product page at x.ai/bot refuses scripted requests, and as of 20 August 2026 the @xai account has never posted about Grok Bot, AI teammates, or x.ai/bot at all. Every capability and pricing claim in circulation traces back to @grok answering users in replies. That is weaker than an announcement, and we have marked it as reported rather than confirmed throughout. The one thing we can state first-hand is the publish test below, which we ran ourselves and recorded.

What it is

The shape, as described by @grok across several replies: each bot gets its own persistent cloud computer. It signs into your tools and apps like a human, typically via browser automation after a human-assisted first login and 2FA. It runs multi-step work continuously, including while your own device is locked. It learns routines from one demonstration and saves them for reuse. Bots can coordinate with each other in group chats with role charters. It completes tasks end to end, with the examples given being file cleanup, email triage, research and shopping, and CRM updates, then checks back for review.

Access is bundled rather than sold separately: SuperGrok Heavy at roughly $300 a month, Cursor Ultra at roughly $200, or Cursor Teams Premium at roughly $120 per seat, with weekly usage included and heavier use billed on by tokens. One account supports around 50 bots. It runs on desktop and iOS, with Android described as coming.

Computer use, not integrations

The important design decision is that Grok Bot reaches the outside world mostly by using a browser, not by calling APIs through an integration layer. No public source shows the bot consuming MCP servers. It opens the web version of an app, signs in, and navigates the interface.

That has a consequence worth stating plainly, because it inverts the usual assumption about agent platforms: a service does not need an integration, a partnership, or an API key exchange to work with Grok Bot. If a person can do it in a browser, the bot can attempt it. The integration question, which normally gates whether a tool is usable inside an agent product, mostly does not apply here.

The trade is reliability and permission. Browser automation is slower and more brittle than an API call, and it inherits whatever signup wall, CAPTCHA or 2FA the service puts in the way. Services that require an account before the first useful action are much harder for it than services that do not.

We asked one to publish a page

On 20 August 2026 we gave a Grok Bot a single instruction: "Create a standard intro page and publish it to stacktree." No plugin was installed, no API key was provided, and no integration exists between the two products.

It replied that it would look up how Stacktree pages work, then reported that it had found the house design guide, wrote a one-page document to that specification, and published it. It returned a live URL, and described the result accurately without being asked: unlisted, viewers need no account, and expiring the following day. It then offered a next step, noting that claiming the page into an account would let it update the same URL in place rather than minting a new one.

That page is still up: the one-pager the bot wrote and published, untouched by us. It was claimed into an account afterwards, which is what removed the 24 hour expiry, so it also demonstrates the second half of the flow. Note what it says about itself near the bottom: "Written as a single HTML document, published with one call to Stacktree. The URL is unguessable; viewers need no account." The bot understood the product it had just used well enough to describe it correctly, from documentation alone.

Our server records confirm the mechanics. The page arrived with no account attached and no payment, as a single 6.6 KB file, with a 24 hour expiry. In other words the bot found and used the anonymous publishing path, which requires no signup, entirely from reading the documentation. That last detail is the whole lesson: the thing that made it work was not an integration, it was that the first useful action did not require an account.

We were not first to notice this. On 11 August 2026 Adam Ludwin posted that here.now "works out of the box" with Grok Bot for the same reason. Two independent services, neither with an integration, both reachable.

The plugin marketplace

Separately from the bot's browser automation, xAI runs a Grok Build plugin marketplace, announced in beta around 11 June 2026 with MongoDB, Vercel, Sentry, Cloudflare and Chrome DevTools. This surface is MCP-based, and it is small and curated: 19 plugins listed as of 20 August 2026.

Listing is an open pull request rather than an application. You add one entry to .grok-plugin/marketplace.json pointing at your public repository and pinning a full 40-character commit SHA, regenerate the component index, run the validator locally, and open the PR for CI and code-owner review. The contributing guide is worth reading closely on two points: it asks that keywords and domains be scoped tightly to your brand, because those fields power the prompt that proactively suggests your plugin, and it warns that a branded plugin sourced from a personal account rather than an official organisation will be questioned.

What this means if you build agent tools

Two things, pulling in opposite directions.

The first is that the integration moat is thinner than it looks. A product that reaches services through a browser can use anything, which means the competitive question stops being "are we integrated" and becomes "is our first useful action reachable without an account". Signup walls, which read as reasonable conversion design to a human product team, are the main thing standing between an agent and your product.

The second is that discoverability replaces integration as the bottleneck. If every service is reachable, the question becomes which one the model reaches for. That is what the plugin marketplace is actually selling, and it is why the keyword scoping in its contributing guide is more interesting than the listing itself.