The Plugin Ecosystem Behind Dots: 4,000+ Apps, One Interface

Every dot's usefulness beyond its own cloud computer depends on one number OpenAI has repeated across its launch materials: connections to more than 4,000 applications through OpenAI's plugin ecosystem. It's the figure that turns "an AI with its own browser" into "an AI that can plausibly touch your CRM, your calendar, your ticketing system, and your design tool without you writing a single integration." That's a genuinely large number, and it's worth being precise about what it does and doesn't tell you.

What the number actually represents

Four thousand-plus is a count of available connections — applications that have some form of plugin integration into OpenAI's ecosystem, the same connective layer that already threads through ChatGPT's broader tool-use features rather than something built exclusively for dots. It's not a count of "apps a dot has been specifically tested against" or "apps where a dot can perform every action a human user could." Plugin depth varies enormously across any integration ecosystem this size — some connections expose a handful of read/write actions, others expose nearly the full surface of the underlying application, and OpenAI's public materials don't break down where any specific app falls on that spectrum.

That distinction matters because it's the same one that determines whether a dot's promise — "connect to more than 4,000 apps" — translates into "your dot can update a Salesforce record" or just "your dot can view a Salesforce record." Both are real capabilities; they're not the same capability, and the headline number doesn't disambiguate between them.

Plugins versus the browser fallback

The plugin catalog is one of two ways a dot reaches the outside world — the other, covered in Every Dot Gets a Computer, is direct browser automation inside the dot's own cloud environment. In practice, these two mechanisms cover different territory. Plugins give a dot structured, presumably more reliable access to apps that have bothered to build the integration; the browser gives it a general-purpose fallback for the much larger universe of websites and tools that haven't. A dot checking a prospect's requirements against internal documentation, one of OpenAI's named sales use cases, might use a plugin to pull structured data from a CRM and browser automation to check a page that has no plugin at all — both inside the same task, without the user necessarily knowing which mechanism handled which step.

How this compares to Grok Bot's very different bet

The plugin-ecosystem approach is also where OpenAI's design philosophy diverges most sharply from xAI's. xAI's own announcement for Grok Bot makes "no APIs, no MCP support required" a headline pitch — Grok Bot signs into tools using the user's own credentials and interacts with them the way a human would, specifically so it can reach tools that have never built, and may never build, a formal integration. We cover that tradeoff directly in No APIs, No MCP, No Problem: The Legacy-Tool Pitch.

Both approaches are legitimate answers to the same underlying constraint: the software world is bigger than any plugin catalog, no matter how large. OpenAI's bet is that a large, structured catalog covers enough of what matters that the remaining gaps can be handled by browser automation. xAI's bet is that credential-based sign-in covers everything a plugin catalog would, without needing the catalog at all — at the cost of handing an agent your actual login. Which bet ages better probably depends on how much of the remaining 4,001st app problem plugin catalogs can realistically close over time, versus how comfortable users and enterprises get with an agent holding real credentials.

A plugin catalog is a promise about breadth. It's not, by itself, a promise about depth — and the gap between the two is where most of the real limitations in "connect to 4,000+ apps" actually live.

What's still genuinely unclear

OpenAI hasn't published a full directory of which specific applications are included, what actions are exposed for each, or how the plugin permission model interacts with Custom Rules at a per-app level — whether, for instance, you can grant a dot read access to a CRM through its plugin while blocking write access, or whether that granularity exists at all. Those are reasonable questions for anyone evaluating dots for a specific stack, and the honest answer, as of this writing, is that the public materials don't go deep enough to say.

What is clear is the strategic logic: a large plugin ecosystem is a genuine moat if maintained, because it turns "does this AI agent work with the tools we already use" from an integration project into a checkbox.

The maintenance problem behind any large catalog

Plugin ecosystems this size also come with a durability question that headline numbers tend to obscure: a catalog of 4,000-plus integrations is only as good as how often each one gets updated when the underlying app changes its own interface or API. Every large plugin or app-store ecosystem eventually accumulates a long tail of stale, half-broken, or abandoned integrations, and there's no public information yet on how OpenAI intends to keep a catalog this size current as the apps behind it evolve independently. That's not a criticism specific to OpenAI — it's a structural challenge for any platform that aggregates third-party integrations at scale — but it's a reasonable thing to watch for rather than assume solved just because the initial number is large.

It's also worth noting what this means for how quickly a dot can become useful for a genuinely new or niche tool. A well-established SaaS product is more likely to already have plugin coverage; a smaller, internal, or industry-specific tool is more likely to fall into the browser-automation fallback, at least initially. That's a reasonable pattern for a launch-week ecosystem, but it does mean the "4,000+ apps" figure will feel more or less impressive depending heavily on how mainstream your particular stack is — a consideration worth keeping in mind before assuming a dot will slot cleanly into whatever specific tools your team actually relies on. Whether OpenAI's 4,000-plus number holds up under actual usage, or whether users find themselves routing through browser automation for more of their real workflows than the headline number implies, is exactly the kind of thing that becomes clearer only with a few months of adoption — which is also true of the more personal layer OpenAI is betting on to keep users invested in a specific dot: giving it a name.

Part 7 of 30 in the ChatGPT Dots series. Previously: Pages, Not Documents: How Space Reimagines Collaborative Work. Next: Naming Your Dot: Personalization as a Trust Mechanism. Browse more builder insights or explore AI skills at aiskill.market.