It might be time to stop worrying about putting AI in your app. The solution might be to put your app in the AI instead.
For software you sell, customers could interact with your product via their own AI assistant. For software your team uses internally, employees could reach company data from the AI tools they already have. The technology is the same in both cases: an API with a Model Context Protocol (MCP) server on top. What changes is whose AI assistant connects to it, and who decides whether to allow it.
When businesses want to add AI features to software, they usually take one of two paths. Either they build an assistant into their own app, or they make their app usable from AI tools people already use, such as ChatGPT and Claude. The second path runs through APIs and MCP. This has made "MCP vs API" a common debate for product leaders.*
For products whose value lives in their data, that second path deserves more attention. That doesn’t mean abandoning traditional UI, though. At Sketch, we often like a "yes, and" approach. In this case, that means you can make your software usable inside AI tools and keep investing in web and mobile.
*Even debating MCP vs API is a little bit tricky from a semantic standpoint. For one, MCP typically sits on top of traditional APIs, rather than replacing them. Also, MCP is technically a type of API designed for AI applications. In this post, ‘API’ means a product’s existing API, like the REST API web and mobile apps already use.
Rather than choosing between building AI into a product or connecting the product to the AI assistants people already use, many large vendors now do both.
For example, HubSpot builds its own Breeze agents and publishes an official Claude connector that lets Claude view and update CRM records. Atlassian builds Rovo into its products and runs a Rovo MCP server that connects to ChatGPT, Claude, Cursor, and VS Code.
The case for the second path largely comes down to context. A user’s assistant already holds their other apps and files, which an in-app chatbot can’t see.
Related reading: Learn how to get started with AI for business.
People already spend a lot of time chatting with AI assistants. In many cases, those chats are becoming the surfaces through which people interact with traditional software tools, too.
One primary purpose of business software is to give people a way to reach structured data, while enforcing authentication, authorization, and rules for how that data changes. Those parts stay. The screen people use to reach them can live in more than one place.
This applies to consumer software, too. Take ride-hailing as an example. A rider wants to get from point A to point B. Instead of opening a mobile app, a request in an AI assistant might be enough. If the assistant had connectors to multiple ridesharing services, it could compare them on price in one request.
For another example, consider parts ordering. A powersports owner asks an AI assistant to order the kit for a scheduled service and approves the purchase. Today the assistant often has to click through a website built for people, which is slow and error-prone. A well-built connector is easier to maintain and easier to use.
The audience is already there. For example, OpenAI says it has 1.2 billion weekly users as of September 2026. When HubSpot launched its ChatGPT connector in June 2025, it said more than 75% of its customers already used ChatGPT (per MarTech).
The leading LLMs are built for this kind of work, too. OpenAI launched apps in ChatGPT, built on MCP, in October 2025. Anthropic added interactive connectors to Claude through MCP Apps in January 2026.
At Sketch, we’ve taken advantage of this with our own software. TimeFront, a tool the team built in-house, can be connected to AI assistants. Approved users can access financial, engagement, and personnel information without leaving Sketch’s Claude Team account. Access is limited by the user’s normal TimeFront permissions.
Some people may come to treat an AI assistant as their main way into software, even over a traditional browser. For now, it's one more surface for reaching software and businesses, alongside web and mobile.
An API lets software request data or actions from another system. MCP (Model Context Protocol) is an open standard that packages those capabilities as tools an AI application can find and use. In other words, an MCP server usually sits on top of an existing API.
An API (application programming interface) has endpoints for specific tasks and information. A ridesharing app’s API might have separate endpoints for fare estimates, driver availability, payment, and ride requests. An MCP server turns these capabilities into tools, and the most useful tools match the jobs people want done. A single “book a ride” tool might call all four endpoints behind the scenes. Anthropic recommends this approach because a short list of focused tools is easier for an AI to choose from and uses less of its context window.
The MCP server publishes a catalog of its tools: what each one does, which inputs it needs, and what comes back. ChatGPT, Claude, and other MCP-compatible applications can read that catalog. Every company designs its API in its own way, which can make them harder to call.
AI applications could already call APIs before MCP, but each app needed its own integration for each one. Ten AI apps and 10 software tools meant up to 100 custom connections. With MCP, each tool builds one server and each AI app builds one client, so the same setup takes 20 builds (10 servers and 10 clients) instead of 100.
Better still, to use Anthropic’s words, “Instead of maintaining separate connectors for each data source, developers can now build against a standard protocol.” For a software company, that means one MCP server can work with ChatGPT, Claude and other MCP-compatible apps.
Because the API is the foundation, an MCP server is only as reliable as the API under it. Postman's 2025 State of the API report found that 89% of developers were already using generative AI, but only 24% were designing APIs with AI agents in mind. That means the remaining 76% of developers weren’t focused on making APIs easier for AI to use.
Two more pieces matter for business leaders:
MCP has moved fast. Anthropic open-sourced it in November 2024 and donated it to the Linux Foundation's new Agentic AI Foundation in December 2025, when there were already more than 10,000 active public MCP servers.
|
|
API |
MCP server |
Command-line tool plus a skill |
|
Who uses it |
Any software: your apps, partner systems, scripts, and AI agents through function calling |
AI applications such as ChatGPT, Claude, Cursor, and VS Code, where the model picks tools as it works |
A person at a terminal, or an AI assistant that can run commands, such as Claude Code |
|
What it offers |
Endpoints, such as "fare estimate" or "payment" |
Tools the AI can use, like “book a ride,” plus optional resources (data) and prompts (templates) |
One command built for one job |
|
How the caller learns what's available |
Documentation, often with a machine-readable OpenAPI description |
The AI application reads the server's list of tools and their descriptions |
The skill spells out which command to run and when |
|
Standardization |
Shared conventions such as REST, with each API designed its own way |
One protocol that many AI applications support |
Custom to whoever built it, though skills follow the open Agent Skills standard |
|
Access control |
Set by each API: keys, OAuth, and others |
Optional in the spec; remote servers that use it follow an OAuth 2.1-based flow |
Runs with the credentials the person set up |
|
Interface in the chat |
None |
Optional, through MCP Apps, where the AI application supports it |
None |
|
Best fit |
Known integrations between systems |
Letting customers use your product from their AI assistant |
A repeated question that needs the same answer every time, and a technical user to create it |
A command-line tool (CLI) is a small program run by typing a command. Paired with something like a Claude skill, a reusable set of instructions Claude applies when it needs them, a CLI gives an AI assistant one narrow, reliable path to an answer.
For example, a custom software engineer keeps needing a relevant Jira story assigned to him and marked in progress as part of his software development workflow. Through the full Atlassian connector, Claude had to work out which board he was on each time. So he wrote a small CLI on the Jira API that returns that one answer, plus a skill that tells Claude to call it. He gets the same answer every time.
This won't suit everyone, and a detailed skill can get similar results without new code. Atlassian's Rovo also handles a lot well once teams learn to use it, which is where Atlassian training helps. Some power users will still want a narrower path. Broad connectors suit open-ended work. Narrow tools suit the questions people ask over and over.
It depends on the product and the people using it. For structured data with defined actions, MCP fits, and an in-app chatbot largely repeats work ChatGPT and Claude can already do. In other cases, like with scattered documents that need curation and guardrails, an in-app custom AI tool is the right call.
As always, the situation and preferences of the user base are important. Users without an AI assistant, or whose employer blocks connectors, won’t get much from an MCP server.
If your company already runs a connected set of internal software and AI tools, it likely makes sense to start with an API and an MCP server. Similarly, for products you sell, an MCP server lets users engage with AI on their own terms, in the accounts they already use.
Simply slapping a chat window into an interface duplicates a lot of what ChatGPT and Claude already do, and maybe not with the same quality. Good product decisions matter more than putting AI on every screen.
Related reading: See the guide to agentic AI development.
With MCP, the user’s AI subscription pays for the AI processing. With an in-app chatbot, you eat the token costs.
Custom AI fits internal teams asking open-ended questions about their own data, products where the visuals are the point, and users who prefer a classic interface, including many executives.
Lowbrow Customs is one example. The motorcycle parts supplier's product knowledge was scattered across PDFs and email drafts, and its support staff needed accurate fitment answers fast. Sketch built an internal AI chat for business use, grounded in that data and designed to flag when it lacks a confident answer. A developer was able to deliver it quickly, and the team started using it right away.
Learn how a custom motorcycle parts company made technical expertise more accessible in the Lowbrow Customs case study.
For internal tools, the downside to exposing your application to AI is mostly data governance. Company data flows through an outside AI provider. Also, employees need licenses to use the software this way, which might not be ideal for every business.
For your public-facing products or products you sell, showing up inside AI chat makes a product easier to compare to alternatives, and hands some control to the AI platforms. Staying out can cost more, because an assistant can only recommend the companies that are there.
Comparison is the first trade-off. An assistant can line up two ride-hailing services on price, which helps customers and weakens brand loyalty. Instead of opening the go-to app, the user interacts with the AI assistant, which weighs options and makes a decision.
Platform dependence is the second. If people mostly reach software through a few dominant AI companies, those companies control the interaction. That's the main reason to keep your own interface, even as you go where your users are.
You can get the best of both worlds by making your software available inside of AI and maintaining UI quality.
No plan fits every product, but here are a few steps to consider:
If you're not sure where your product stands, an AI readiness assessment is a useful first step.
An API lets software request data or actions from another system. MCP turns those capabilities into tools an AI application can find and use, usually on top of an existing API.
No. Most MCP servers call an existing API to do the work, so the API stays the foundation.
It depends. For software you sell, an API and an MCP server let customers use your product from the AI assistants they already have. An in-app chatbot fits better when visuals matter or when users prefer a classic interface.
For internal tools, an MCP server suits structured data, like CRM records or project tickets. A custom chatbot suits knowledge that’s scattered across documents and needs curation. Custom chatbots also make sense when data has to stay in your own cloud environment, since a custom tool can use AI models hosted there.
Yes. Through MCP Apps, an MCP server can show forms, dashboards, and confirmation steps inside the chat, in AI applications that support it.
MCP's specification defines an OAuth 2.1-based authorization flow for remote servers and notes that the protocol can't enforce security on its own. Security concerns to consider include prompt injection, malicious tool descriptions (“tool poisoning”), and over-broad permissions. The spec itself says tool descriptions should be treated as untrusted and that tools can run arbitrary code.
In Postman's 2025 survey, 51% of developers named unauthorized or excessive API calls from AI agents a top security worry. Safety depends on the consent flows and permissions you build.
Sketch Development Services builds custom software, APIs, and custom AI solutions with a 100% US-based team. Sketch won the Clutch Summer 2026 Global Award for API development, which speaks to years of enterprise experience with APIs. The team now builds MCP servers, too. If you want customers to use your product inside ChatGPT and Claude while you keep your own interface, talk to Sketch about whether it's a fit. The first couple of consultations are free.