Musk Hung a Sidebar Inside Microsoft's Living Room

On Tuesday, July 21, a new listing appeared on the Microsoft marketplace. Name: Grok by SpaceXAI for Outlook. Product ID: wa200011330. On the surface it is the most boring thing in software — an Office extension. You click install, a button shows up in the Outlook ribbon, a panel slides open on the right, and you start typing at it.

What makes it not boring is the real estate. That panel opens a few pixels away from Microsoft Copilot, the assistant Microsoft has spent billions wiring into every corner of Office. So here is the shape of it: Musk's company used Microsoft's app store, passed Microsoft's certification, called Microsoft's APIs, and planted a Microsoft competitor inside a Microsoft product.

That single image is the whole story. It reads less like an invasion and more like a tenant filing paperwork. To move into someone else's building you need the landlord's sign-off, and SpaceXAI got it. Microsoft has said nothing publicly about the listing — no comment from Redmond turns up in any search.

One more framing note before we go deeper: this shipped on July 21, which makes it ten-day-old news as of today, July 31. A lot happened inside those ten days. SpaceXAI ran the same play against Google Workspace the following week, and the Grok Build data leak that broke nine days before the Outlook launch still has no resolution. Read the Outlook add-in on its own and you miss half the plot.

The Rocket Company That Swallowed the AI Company

Let us clear up the name first, because it matters. The company that shipped this add-in is no longer called xAI. SpaceX absorbed xAI in February 2026, and in May the combined entity rebranded as SpaceXAI. It was an all-stock triangular merger that valued SpaceX at $1 trillion and xAI at $250 billion.

Then, on June 12, that merged entity listed on Nasdaq under the ticker SPCX. It offered 555.56 million Class A shares at $135 apiece, targeting roughly $75 billion in proceeds, with an expected market cap around $1.75 trillion at the open. That is the largest IPO in the history of public markets. Buy one share of SPCX today and you are buying rockets, satellite internet, frontier AI models, and a social network in a single wrapper.

Why does the corporate structure matter for an email plugin? Because Grok is now a product that has to defend a public company's quarterly numbers. The SpaceXAI segment holds the X platform, the Grok models, and a fast-growing pile of AI compute, and it is reportedly burning on the order of $1 billion a month. Covering AI losses with launch revenue and Starlink subscriptions is a founder's prerogative when you are private. Once you are public, it becomes a line item you explain every ninety days.

Which is why the Office add-in blitz landing right now is not a coincidence. PowerPoint came first on June 16, then Word, then Excel, then Outlook on July 21 — all four major Microsoft 365 apps covered in roughly five weeks. On July 21 the company also appeared on the Google Workspace Marketplace, formalized with an announcement on July 24 that extended coverage to Docs, Sheets, and Slides. This is a company trying to stop selling a model and start living inside a workflow.

One last detail. The add-ins run on Grok 4.5, the generation released this month. Some coverage cites specific parameter counts for that model; those figures are not confirmed by any SpaceXAI primary source, so they are staying out of this piece. What is confirmed is only that the Outlook add-in sits on the Grok 4.5 generation.

What It Actually Does — and What It Actually Asks For

Start with the features. Per SpaceXAI's announcement and the marketplace listing, the Outlook add-in does four things. It summarizes long threads and pulls out decisions, action items, and messages you have not answered yet. It drafts replies that imitate your normal writing voice. It cleans your inbox — archiving finished conversations, filtering noise, flagging what needs a response. And it reads attachments and searches the web and X without making you leave Outlook.

There is one genuinely good safety design here. Anything Grok writes stays a draft until you personally hit send. No autonomous sending. In email AI that is a meaningful line — a bad code suggestion can be deleted, a bad email cannot be unsent.

Then you look at the permissions and the mood changes. SpaceXAI's own developer documentation spells out the scopes: the Outlook mail connector requests Mail.ReadWrite (manage mail and drafts), Mail.Send (send messages), User.Read (read your profile), and offline_access (persistent access). Hook up the calendar and Calendars.ReadWrite gets added. Read, write, and send all clear in a single consent click. "Does not auto-send" and "does not hold send authority" are two very different sentences, and only the first one is true here.

To be fair, the blast radius is bounded. Every one of those is a delegated permission, meaning Grok sees only the signed-in user's own mailbox and calendar. These are not application permissions that let a service sweep an entire tenant's mail. Organizations on Azure AD may require admin consent first, and administrators can approve or block the app from the Enterprise Applications panel. Users can sever the connection any time from the connector page or Microsoft's app management portal.

On data handling, SpaceXAI's documentation makes two flat claims: "We do not train on your data," and "Nothing is stored." The stated model is real-time retrieval with no retention. How much weight that deserves is a question the skeptical section below takes up.

Pricing is where it gets murky. The add-in itself installs free from the marketplace. Using it requires a paid subscription. SpaceXAI says it is available to "all paid X and SuperGrok subscribers," and read literally that opens the door at $8 a month with X Premium. But other coverage reports that individual users need the $30-a-month SuperGrok tier, with businesses buying separate Business or Enterprise plans. For the Excel add-in, The Register reported it works on SuperGrok, Heavy, Business, and Enterprise plans. That is a real contradiction and I am leaving it visible rather than papering over it — whether the true floor is $8 or $30 cannot be settled from primary sources alone.

Item Grok for Outlook Microsoft 365 Copilot Gemini in Workspace Superhuman Mail
Launch / current status 2026-07-21 Launched 2023, current Plans consolidated 2025-03 Current
Add-in install cost Free N/A (built in) N/A (built in) Separate client
Minimum monthly spend to actually use X Premium $8 or SuperGrok $30 (sources conflict) $30/user add-on ($21 SMB annual) Included in Workspace price (Business Standard ~$14–18) Starter $25 annual, $30 monthly
Higher tier Heavy $300 Copilot Chat free for existing M365 subscribers Separate add-on discontinued Business ~$33 annual, $40 monthly
Mail permissions Mail.ReadWrite + Mail.Send + offline_access Platform owner's own permissions Platform owner's own permissions Own mail client permissions
Auto-send No (stays a draft) Draft-first Draft-first Auto Draft (higher tiers)
Office coverage Outlook, Word, Excel, PowerPoint All of Office + Teams Gmail, Docs, Sheets, Slides, Meet Email only

The row that should stop you is the third one. Microsoft sells Copilot as a $30 per-seat add-on. SpaceXAI is bundling into a subscription plenty of people already pay for other reasons. If you already have X Premium, the marginal cost of adding an AI assistant to your inbox is zero. This is not a feature fight. It is a billing-structure fight, and those are the ones that actually move seats.

Who Walks Away With What

SpaceXAI gets two things. The first is presence. Until now Grok lived on the X timeline and at grok.com — places you have to deliberately go when you decide "I want to use AI." An Outlook sidebar inverts that. Grok now sits inside a window that is already open on your screen for hours a day, and it gets asked things it would never have been asked otherwise. The second is data adjacency, though since the company explicitly states it neither stores nor trains on this data, that benefit stays theoretical on the company's own word.

Microsoft gets something too, which is less obvious. It looks like pure loss — a rival assistant one panel over from its own. But the existence of competitor AI add-ins in its marketplace is a regulatory shield. Microsoft has spent years under European antitrust pressure over Office bundling, and "we do not block third-party AI assistants, here is the receipt" is worth real money in that context. There is also a lock-in argument: every add-in makes Office stickier, not weaker. Nobody leaves Outlook to use Grok. They stay in Outlook because Grok is there.

Users get choice and a price break. Until now, putting AI inside Outlook meant buying a Copilot license, which in practice meant convincing your employer to buy it. Now an individual can install AI into their own inbox with their own subscription and zero procurement conversations. Which is exactly why the next paragraph exists.

Corporate IT gets, honestly, very little — mostly a new item on the risk register. Tenants that require admin consent are in decent shape. Tenants that leave user consent open now have a situation where any employee can personally hand read, write, and send authority over a corporate mailbox to an outside company, in three clicks, with no ticket filed. This is not a novel category of problem; it is shadow IT wearing a new hat. What is different is the payload. The thing being delegated is the company's entire email correspondence.

The Companies That Won on Someone Else's Platform, and the Ones That Died There

The canonical success story for riding a platform is Grammarly. It shipped a Microsoft Word add-in around 2016, camped inside Office, and survived Microsoft repeatedly strengthening Word's own grammar checking and Editor features — growing to hundreds of millions in revenue anyway. The ending is symbolically perfect: in July 2025 Grammarly acquired the email client Superhuman, and in October 2025 renamed the parent company Superhuman outright. Superhuman today is both an email app and the holding brand over Grammarly and Coda.

Zoom rhymes with that. In early 2020, the Zoom add-in for Outlook got installed at pandemic speed, and then Microsoft bundled Teams into Office at effectively zero marginal cost and pressed hard. Zoom did not die. The lesson is unglamorous but real: when the platform owner gives the same feature away free, a clearly better product still gets kept alongside it. Users will run two things.

The other column is longer. Mailbox, the email app Dropbox bought in 2013, was shut down in 2015. Sunrise, the calendar app Microsoft bought in 2015, was killed in 2016 with its features absorbed into Outlook. Building an excellent mail or calendar client and building a durable business out of one turned out to be different problems entirely.

The scariest precedent is Slack. Slack shipped before Teams, and by most accounts had the better product for years. Microsoft bundled Teams into Office 365 and ground it down anyway. Slack filed an EU antitrust complaint in 2020 and sold to Salesforce in 2021. When a platform owner genuinely plays the bundling card, being better is not sufficient. Whether Grok for Outlook ends up on the Grammarly side of that ledger or the Slack side is, frankly, far too early to call.

The Cards Microsoft Holds, and the Different Card Google Played

Microsoft has three moves available. The first is to do nothing at all. Copilot already ships everywhere in Office, and in Excel it goes as deep as a cell-level COPILOT function you can call from a formula. A third-party add-in takes prompts in a sidebar; Copilot is attached to the document object model itself. That depth gap is not something the add-in API can close.

The second move is API and marketplace policy. Microsoft writes the permission model and the certification bar for Outlook add-ins. It could tighten consent requirements, force admin approval on particular scope combinations, or change listing rules. Every one of those levers is in Redmond's hand. But pulling any of them relights the antitrust fire, so the incentive to actually use them is weak.

The third move is price, and Microsoft already made it. On December 1, 2025, it cut SMB Copilot pricing permanently from $30 a month to $21 on annual commitment and $25 month-to-month, and ran an $18-a-month small-business promotion through June 30, 2026. It also opened Copilot Chat to existing Microsoft 365 subscribers at no extra charge. In other words, Microsoft lowered its price wall months before Grok showed up at it. Whatever this add-in does, it is arriving at an already-fortified position.

Google chose a different route entirely. In March 2025 it eliminated the separate Gemini for Workspace add-on and folded Gemini into paid Workspace plans, raising base prices by roughly $2–4 per user instead. Today a Business Standard customer gets Gemini already switched on across Gmail, Docs, Sheets, Slides, and Meet. With no separate AI line item, there is less crevice for a "free third-party add-in" to wedge into. SpaceXAI went in anyway — on July 24 it shipped a free Grok add-on for Google Workspace covering Docs, Sheets, and Slides.

Superhuman Mail attacks from another angle again. It is not an add-in; it is a whole replacement email client at roughly $25 a month on annual Starter billing and about $33 on annual Business. It costs far more than Grok and sells anyway, because what it is selling is a speed-and-keyboard-shortcut client experience, not an AI feature. Grok's add-in bolts onto Outlook's existing UI, so it simply does not compete on that axis.

OpenAI is quietly walking the same ground from the opposite direction. ChatGPT already supports connectors for Gmail, Outlook, and calendars — but the vector is reversed. Grok put AI inside your email. ChatGPT pulled your email inside the AI. Which of those matches how people actually work is genuinely unsettled.

Now About What Happened Nine Days Before This Launch

Here is the uncomfortable part. On July 12 — nine days before the Outlook add-in shipped — an AI safety researcher published a packet-level capture of what the Grok Build CLI was sending over the wire. The finding: Grok Build was not merely uploading the files a developer asked it to read. It was uploading entire tracked git repositories, complete commit history and any secrets baked into that history included, to a Google Cloud Storage bucket. The volume worked out to roughly 27,800 times the data the actual task required.

The privacy toggle makes it worse. xAI had marketed the tool with an explicit statement that your codebase is not transmitted to xAI servers during a session, and the interface gave users a switch for exactly that. The switch controlled nothing. What actually governed the behavior was a server-side flag users could not see, and the uploads only stopped once developers flipped disable_codebase_upload to true.

The analysis hit the Hacker News front page on July 14, and Musk publicly committed to deleting all user data collected up to that point. As of today, July 31, five things still have not been disclosed: how many users were affected, how much data was ingested since the beta opened in May, how an individual developer can verify their own data was deleted, when deletion completed, and any third-party confirmation that deletion happened at all. Zero for five.

Now go back and reread the Outlook connector documentation's "We do not train on your data" and "Nothing is stored" with that context loaded. Both sentences are clear and may well have been written in complete good faith. The problem is that nine days earlier, at the same company, a different product's on-screen privacy indicator diverged from its actual network behavior — and it took an outside researcher's packet capture, not the company, to surface it. Policy language that nobody audits is just language.

And this time the payload is not source code. It is email. A corporate mailbox holds contracts, personnel discussions, unreleased financials, customer personal data, and privileged legal correspondence. A leaked repository hurts a company. A leaked mailbox gets a company sued. If you run a tenant with an admin consent gate, the minimum bar before approving this add-in is reading the contractual data-processing terms and confirming what audit rights, if any, you have.

One more, less dramatic skepticism: when The Register actually tested the Excel add-in on July 22, requests repeatedly failed against usage limits. And Microsoft itself puts a warning on the third-party add-in install screen: using this app allows it to read and change your document and to send data over the internet.

What Changes, Depending on Who You Are

If you run IT or security at a company — this is a today item, not a someday item. If your tenant has user consent enabled, employees may already have connected this. Go to Azure AD Enterprise Applications, query consent status for the SpaceXAI app, and count how many accounts have delegated Mail.ReadWrite and Mail.Send. Whether to block it is the second decision; knowing the number is the first, and you cannot make the second without it. Simply flipping the tenant back to required admin consent closes most of the exposure by itself.

If you build software — the interesting artifact is not the add-in, it is the proof. Someone publicly demonstrated that the Office add-in API plus Microsoft Graph delegated permissions is enough to construct a mail assistant roughly comparable to Copilot's sidebar behavior. The barrier to entry in this category was never the model; it was distribution and trust. SpaceXAI solved distribution with a marketplace listing. Trust is still open, and the Grok Build incident made it more open, not less.

If you own SPCX — treat this as a narrative event rather than a revenue event. It signals that a company two months into being public is trying to move its AI segment from selling model access to selling subscription workflow, which is the higher-multiple business. But there is not a single disclosed figure showing how many new X or SuperGrok subscriptions the Outlook add-in has converted. The first real test is next quarter's report and whether SpaceXAI segment subscription metrics get broken out with enough granularity to see it.

If you just get too much email — if you already pay for X Premium or SuperGrok, trying this costs you nothing incremental, and the draft-only design is a reasonable guardrail. But do not attach it to your work account. Run it on personal mail first, and if you want it on corporate mail, ask IT before you click consent, not after. One habit risk worth naming: trusting Grok's summary and skipping the original thread. On the threads that matter most, the thing a summary drops is usually the conditional clause that decides everything.

🥄 Three Things You’re Probably Wondering

— So what does this mean for me? If your company runs on Outlook, it means something even if you never install it. The moment a colleague connects this add-in, your emails sitting in their inbox fall inside Grok's processing scope too. It looks like an individual choice and functions like an organizational data-boundary decision.

— Does it really store nothing, like xAI says? The official documentation is unambiguous: no training, no storage. But nine days before this launch, an outside researcher showed that on another xAI product the on-screen privacy control and the actual transmission did not match, and the company still has not produced third-party verification of its cleanup. There is no evidence the Outlook documentation is false — but deciding whether documentation alone is enough to open your corporate mailbox is a judgment call each org has to make itself.

— Will Microsoft actually tolerate this? For now, probably yes. The third-party add-in ecosystem is an Office asset, and evicting a competitor invites antitrust scrutiny Microsoft does not want. Instead it has already cut Copilot pricing and made Copilot Chat free for existing subscribers. Squeezing on price and integration depth beats blocking outright — though with no official comment from Redmond, calling it is still premature.

Further Reading

Numbers and criteria are as of announcement and may change. Investment calls are yours to make!