In the first article of this series, we mapped what Grok Bot is — from public xAI docs — and what Orauria already ships for product images and video. This piece is an architecture note for a tech lead or technical founder who will attach the two systems or hire a team to do it.
It is not a partnership announcement, and it is not a click-by-click tutorial. Grok Bot facts come from public pages on x.ai/bot and docs.x.ai, checked on 11 November 2026 (Asia/Saigon). Orauria facts come from Orauria’s own site, especially the MCP page. We describe three layers that fit those documents. We do not claim an official partnership with xAI. We do not claim that Orauria’s MCP server has been verified end-to-end inside Grok Bot. A connection tutorial will be published only after that test.
Key takeaways. Treat the connection as three layers: a contracted MCP tool path, a browser path on the Bot’s cloud computer, and an orchestration layer of skills, routines, and multiple Bots. MCP is the intended default because a tool call is easier to repeat and test than a click path through a web UI. Grok Bot documents a custom Remote HTTPS connector — server URL, optional HTTP headers — or a Command type. Orauria documents a Bearer API key and six tools that bill a Studio credit wallet. Whether Grok Bot will accept and store Orauria’s Authorization: Bearer header has not been independently verified here. Workflow Builder, Fashion Poses, upscale, and Chrome extensions are not MCP tools. And because the cloud computer is shared across every Bot on an account, API keys belong in a connector header or a secret card — never in ordinary chat, and never in /workspace.

Three layers: MCP, the browser, and orchestration
A useful architecture starts by naming the layers, not the vendor logos. When a merchandising brief arrives, which contract does the Bot actually call — and what happens when that contract is not enough?
Layer 1 is the tool path. Model Context Protocol gives a Bot a small, explicit set of functions it can call over a documented interface. On Orauria today, that set is public: generate_image, generate_video, chat, list_models, get_wallet, and get_task. A call has a typed result you can inspect — an image URL, a job id, a credit balance — which is what you want if a routine or a specialist Bot will run the same step tomorrow.
Layer 2 is the browser path. Every Bot on a Grok Bot account shares a persistent cloud computer with a browser, a filesystem, and a terminal. When a job needs a visual UI that MCP does not expose — a node canvas in Workflow Builder, a Fashion Poses board, a click-through in Studio — the Bot can open the public site in that browser, the same way it would open any other page. This path is available today. It is also harder to test, easier to break when the UI moves, and a poorer home for a standing routine.
Layer 3 is orchestration. Grok Bot’s skills, routines, and multi-bot chats sit above both paths. One Bot can write a brief; another can call image tools or watch a job; a routine can fire on a schedule. The collaboration model is documented. Wiring it to Orauria still waits on a verified connector, and on skills we will only publish after they run for real.
If you only remember one sentence from the opening guide, let it be this: MCP is the intended default; the browser is the escape hatch; skills and routines are how you stop doing the same handoff by hand.

Why MCP is the preferred path
There is a reason we keep calling MCP the default, even though the browser path is the one you can use this afternoon. The difference is whether the work can survive contact with a calendar.
A UI click is a demonstration. You watch the Bot open Studio, choose a model, wait, and download a frame. That is useful for a one-off, and it is the honest answer for any Studio surface that has no MCP tool. It is a weak foundation for a routine that should run every weekday at 09:00, or for a specialist Bot that should generate a packshot set without you watching the computer preview. Layouts change. Buttons move. The Bot has to see, infer, and click again. You can recover, but you cannot write a tight skill around that recovery.
A tool call is a contract. get_wallet returns a balance. list_models returns the current catalog. generate_image, when it succeeds, returns a URL you can file or hand to the next step. get_task lets you poll an asynchronous job — video especially — until it finishes or fails in a way the Bot can report. Success and failure are data, not a screenshot of a spinner. That is also why MCP is the better home for a team of Bots: a merchandising Bot and a retouching Bot can share named tools without sharing a click script or the same brief.
The browser path is not second-class. It is the right tool for the remainder: Workflow Builder’s node canvas, any Fashion Poses or upscale flow, and any review that is simply faster in a visual UI. The rule is narrow: prefer a callable contract when one exists; open the UI when it does not. Do not pretend a missing contract is already there because the page looks impressive.

Custom connectors, as Grok Bot documents them
Grok Bot’s public computer-and-apps documentation describes how to add a custom connector in the desktop app. You choose Add a custom connector at the bottom of Connect apps, give it a name, and pick a type. Remote HTTPS takes a server URL, with optional HTTP headers and an OAuth Client ID — a Client Secret field appears once you enter one. Command takes a command and its arguments. You then save the plugin. The same documentation says you can also ask a Bot to add an MCP server that is not in the catalog, by URL, or by a command that runs on its computer.
That shape is what we mean by “documented design.” It is the interface Grok Bot already publishes for third-party MCP servers. It is not, by itself, a verified recipe for Orauria.
Orauria’s side of the contract is also public, and narrower than a casual reader might hope. The MCP page documents how to connect Cursor, Claude, and Codex today. Auth is a Bearer token: an sk-… API key created in your Orauria profile, sent as Authorization: Bearer. Tools bill the same Studio credit wallet you see on the pricing page. The technical endpoint is written in places as http://orauria.com/api/mcp; HTTPS should be preferred, and that exact URL has not been independently confirmed here. Documented clients, as of the date on this article, do not include Grok Bot.
Put the two documents side by side and a plausible mapping appears: a custom Remote HTTPS connector would carry Orauria’s server URL and an Authorization header, and a Bot would call list_models or get_wallet without pasting a key into the transcript. Two things remain unverified. First, whether Grok Bot will accept, store, and actually send that Bearer header on each tool call. Second, whether the HTTPS endpoint we would prefer is the one that should be entered, given the http:// form still visible in docs. A later tutorial will record those results with real, redacted screenshots. Until then, do not treat the mapping as a setup guide, and do not paste a key into chat to “just try it.”

The current boundary of Orauria’s MCP tools
Architecture goes wrong when it draws boxes around products that do not yet have an interface. Orauria is a creative workspace for product marketing — Studio, Workflow Builder, specialized apps, and a credit wallet. Only a slice of that surface is callable through MCP.
The six tools have clear jobs. list_models tells a Bot which image, video, or chat models it can use. get_wallet is the gate before you spend: read a balance, and stop cleanly if the wallet is empty. generate_image and generate_video create the work and bill that same wallet at Studio prices. get_task is how you wait on a long job — particularly video — without guessing. chat calls Orauria’s billed chat models through the same server. That is a complete, if thin, pipeline for “brief in, media URL out,” as long as the brief fits those tools.
What is not in the list matters as much. There is no MCP tool to run a Workflow Builder graph, and none for Fashion Poses, upscale, or the Chrome extensions. Those are real Studio capabilities, and they remain Layer 2: the Bot can open them in its cloud browser, or a person can open them in their own. They are not contracts you should put in a skill and promise will fire unattended.
If you are designing a system, or briefing a vendor to extend one, keep this boundary visible. The honest diagram is smaller than a single box labeled “Orauria.” MCP covers generation, model discovery, wallet checks, job polling, and chat. Everything else is UI, a future tool, or a custom server someone has to build — which is exactly the kind of work a custom AI team can take on, but not something this architecture should pretend already ships.

A shared cloud computer, and the credential risk
Grok Bot’s documentation is blunt about the computer. Every Bot on your account uses the same one. Browser cookies and signed-in sessions are shared. Files are visible to every Bot. Command-line credentials are shared. One Bot can continue from work another Bot saved. The computer is assigned to your user account, not to an individual Bot. Isolation between users is strict; isolation between your own Bots is not.
That design is what makes handoffs between Bots possible. It is also a credential hazard if you treat the computer like a private laptop. A store login you complete so one Bot can download a reference pack is available to the next Bot that opens a browser. A file in /workspace is not a vault. An API key pasted into a prompt, saved in a scratch file, or typed into a page the Bot is watching can leak into a transcript, a log, or another role’s context.
The safer pattern follows both products’ public advice. Orauria’s key should live as a connector header — if and when that path is verified — or be entered through Grok Bot’s secret card or Secure Form when a human has to type it. It should not be pasted into ordinary chat, and it should not be written to /workspace. If you sign into orauria.com in the cloud browser for Layer 2 work, remember that the session is account-wide: do not leave a privilege on that computer that a less trusted Bot should never have. We will return to vaults and key rotation later in this series. The point already belongs on the architecture diagram: a shared computer is a shared trust boundary.
Design tradeoffs, and what to do next
Three decisions follow from the documents we have, and they are enough to brief a build.
First: MCP-first. If a step has a documented tool — generate, poll, list models, check the wallet — that is the path a skill or routine should call. It is the path you can test, and the path a specialist Bot can share without inheriting a click script.
Second: use the UI when the tool does not exist. Workflow Builder, Fashion Poses, upscale, and the Chrome extensions stay on Layer 2 until they have contracts of their own. The cloud browser is a valid, documented way to reach them. It is not a substitute for a missing tool in an unattended routine.
Third: do not inflate the diagram. Do not label Orauria as an official Grok Bot app. Do not assume the Bearer header already works. Do not hang a full merchandising OS on six tools. The thinness is a feature. A small, billing-aware MCP is something you can reason about, hire against, or extend without pretending the rest of Studio is already an API.
To see how that MCP is documented for the clients it already supports, start at orauria.com/mcp. If you need a connector, a workflow, or a pipeline that does not yet exist as a tool, send a brief to Orauria’s custom AI team. A step-by-step tutorial for attaching Orauria to Grok Bot will come only after we have run that connection ourselves. Until then, this architecture is the map: three layers, a clear boundary, and no keys in the wrong place.







