MCP Distribution
ChatGPT, Claude, Microsoft Copilot, and Gemini Enterprise all reach external systems through the same protocol. That means the expensive engineering happens once, and reaching the next assistant is a distribution decision rather than another integration project.
The premise
The Model Context Protocol is an open standard. An assistant that supports it calls your server, and your server decides what it is allowed to see and do. Because the major assistants have converged on the same protocol, the work that costs real money is portable:
What does not carry over is the listing, the review, and some of the presentation. Those are per-platform, and they are the smaller half of the job.
The surfaces
Surface: Apps and connectors inside ChatGPT, with a public directory
Why it matters: The largest consumer assistant audience, and the only one of these surfaces with a public directory big enough to study. It is where most companies see customer demand first.
What to watch: Public listing means review: published policies, an authentication model, and predictable tool behaviour. Interactive app components are a distinct piece of work on top of the MCP server.
Surface: Connectors for Claude, including Claude for Enterprise deployments
Why it matters: MCP originated here, and Claude is disproportionately represented inside engineering and knowledge-work teams. For B2B data products, the audience quality often beats the raw audience size.
What to watch: Enterprise rollouts frequently want the connector pointed at internal systems behind the customer’s own identity provider, which changes the permission model more than it changes the server.
Surface: Agent and connector surfaces across Microsoft 365
Why it matters: Distribution into companies that have already standardised on Microsoft 365. The assistant sits next to the documents, mail, and meetings your customer already works in.
What to watch: Tenant administration and governance are the gating factors more often than engineering. Plan for how a customer’s IT organisation approves and deploys the thing, not just how it works.
Surface: Enterprise assistant deployments on Google Cloud
Why it matters: Relevant where the customer’s data already sits in Google Cloud and the assistant is being rolled out as an enterprise programme rather than adopted bottom-up.
What to watch: These are typically contracted enterprise deployments, so the work looks more like an implementation programme than a marketplace launch.
Evidence
ChatGPT is the only one of these surfaces with a public directory large enough to analyse, so it is the best available proxy for how far the ecosystem has actually got. We catalogued all 2,289 listed apps in August 2026.
95.6% of the 2,007 companies with a listing have exactly one app, and only 6.7% define named skills. Read across to the other platforms and the implication is the same: almost nobody has built the portable asset yet, let alone distributed it more than once.
The full analysis is in The State of ChatGPT Apps, and the catalogue itself is in the AI Connectors Directory.
Sequence
The right first platform is wherever your customers are already trying to get answers about you. Raw user numbers matter less than whether the people asking are the people who buy.
Largest audience, a public directory, and the clearest discovery path for someone who does not already know your brand.
Smaller audience, heavily weighted toward the engineering and analyst teams that many B2B data products are actually sold to.
If your customer runs Microsoft 365 or Google Cloud end to end, the assistant their staff will use is usually the one already deployed to them, and the deciding factor is tenant governance rather than product preference.
A first connector typically takes eight to twelve weeks. Adding a second surface afterwards is a fraction of that, because the server, the tools, and the permission model already exist.
FAQ
No, and that is the main argument for building on MCP. The Model Context Protocol is a shared standard, so one well-built server can serve ChatGPT, Claude, Microsoft Copilot, and other MCP clients. The expensive work is the tool design, the permission model, and the integration with your systems, and that part is done once.
Three things, mostly. The listing and review process, which is per-platform and varies in strictness. The presentation layer, since platforms differ in what interactive components they support. And the distribution model, which ranges from a public consumer directory to an enterprise tenant deployment approved by a customer’s IT organisation.
Usually wherever your customers are already asking. For consumer-facing and SMB products that is typically ChatGPT, which has the largest audience and a public directory. For products sold into engineering or knowledge-work teams, Claude often converts better. For anything sold into large Microsoft-standardised enterprises, Copilot is the shortest path to the user. Start with one, prove the tool set, then distribute.
Materially less than the first, because the server, the tools, and the permission model already exist. What remains is the platform-specific listing, review, and any presentation work. The first connector is the eight-to-twelve-week project; subsequent surfaces are usually a fraction of that.
No. Plenty of connectors are built for internal teams or for a company’s own authenticated customers and are never published in a directory. Public listing is a distribution decision that you can make after the server exists.
The Model Context Protocol is an open standard that lets an AI assistant reach external systems through a defined set of tools. Instead of the assistant guessing from training data, it calls your server, which decides what it is allowed to see and do. It is the same protocol across the major assistants, which is what makes a single server portable.
In a strategy session we work out the best first use case, which data to expose, whether your current APIs can support it, and which assistants to prioritise.