Stop Building AI Into Your App. Let the AI Come to You.
If your roadmap says “add AI”, you’re probably planning a chatbot, a copilot sidebar or a “summarise” button. For most apps there’s a cheaper, more effective option: make your app available to the AI agents your users already use, through MCP, the Model Context Protocol.
This post compares MCP with embedded AI, explains when each makes sense, and shows what it takes to make your app work with ChatGPT, Claude and other AI apps.
What is MCP?
MCP is an open standard, introduced by Anthropic in late 2024, that lets AI agents discover and use outside apps. An MCP server publishes a set of actions, like search_rooms, get_booking or create_reservation, each with a name, a description and the inputs it takes. Any AI app that supports MCP can connect, read those actions and use them for the person it’s working for.
In practice, MCP turns your app into something the user’s agent can use, instead of one more place the user has to go.
Why embedded AI features underperform
Teams that ship an in-app assistant tend to run into the same problems:
- Low adoption. People already have a favourite AI assistant, and they rarely switch to yours.
- Narrow context. Your assistant only knows your app. The user’s own agent knows their calendar, inbox, files and other tools.
- Model drift. The best models improve every few months, and you’re maintaining prompts against whichever one you integrated last.
- Inference cost. Every conversation is a token bill that grows with usage, not revenue.
- Support burden. When your assistant gets something wrong, that’s your problem.
Your users already chose their AI
People start tasks in ChatGPT and Claude. Developers work in Cursor. More and more people use assistants like Grok and Meta AI every day. When they want something done in your product, they want to ask their agent. Your embedded assistant competes with that, and it usually loses.
MCP vs embedded AI, side by side
| Embedded AI (chatbot or copilot) | MCP | |
|---|---|---|
| Where users ask | Inside your app only | ChatGPT, Claude, Cursor and other AI apps that support MCP |
| Who pays for the AI | You | The user’s AI provider |
| Model upgrades | You re-integrate | Automatic, through the user’s agent |
| Context from other apps | None | Whatever the user’s agent already knows |
| What you build | Chat UI, prompts, guardrails | Actions, permissions, sign-in |
| Reach | Users who open your app | Every AI app your users bring |
How to make your app work with ChatGPT, Claude and other AI agents
- Start from your existing API. Your OpenAPI file is a good starting point, but it isn’t enough on its own (see what breaks when real agents use real APIs).
- Offer task-level actions. Expose what people actually ask for, like “book a class”, rather than every endpoint.
- Describe them clearly. Agents choose actions by reading their descriptions and inputs.
- Let each user sign in. The agent should act as a specific person, with that person’s access and nothing more.
- Be careful with changes. Payments, cancellations and deletions should wait for the user’s OK (see MCP security for write access).
- Host and watch it. Agents expect an always-on connector, and you’ll want to see how it’s used.
When embedded AI still makes sense
Keep building AI into your product if:
- AI is the product (writing, design, code).
- The work is deeply visual and doesn’t fit into actions.
- You need a specialised model general agents don’t offer.
If your AI feature is a thin chat layer over your own API, MCP reaches more people for less.
Frequently asked questions
Do I have to remove my existing AI features? No. An MCP connector sits alongside your app. Many teams keep a small in-app feature and add MCP for reach.
Which AI apps support MCP? Claude, ChatGPT and Cursor support MCP connectors, and more AI apps are adding support. The details differ between apps and change quickly.
Is MCP secure enough for business data? MCP is a protocol. Security depends on how sign-in, per-call permission checks, approvals and isolation are done. That’s the hard part, and the part worth getting right.
How long does it take? A basic MCP server can take a day to build. A production one, with each user’s own sign-in, approvals, hosting and monitoring, takes much longer unless you use a platform that does it for you.
Put your app inside every AI agent
Gephyr turns your existing API into a hosted, secure MCP connector. Paste a link to your API docs, and Gephyr AI suggests a short list of clear actions. You choose what agents can read, change or must ask first about, connect the login your users already have, and publish one link. You write no code, and building and testing are free.