MCP Server Development
Make Your Existing Website AI-First
A custom MCP server so Claude, ChatGPT and other assistants can use your product as a tool, listed where their users find it, plus the groundwork that lets AI search read and cite your site.
More and more people start in an AI assistant instead of a search box. When they ask it to find an apartment, book a service or look something up, the assistant either reaches your product through a tool it can call, or it reads whatever it can scrape and moves on. I take an existing website or product and add that tool layer: an MCP server built on your real data and actions, WebMCP for agents working inside the browser, listings in the directories where Claude and ChatGPT users find connectors, and the crawler, llms.txt and structured-data work that decides whether AI search can read and cite you. Nothing gets rebuilt. It sits on top of the site you already have.
Start a project →What you get
What an AI-first website needs
Custom MCP Server
A remote MCP server on top of your existing database or API, exposing the actions that matter: search, look up, check availability, request a quote, book. Every tool gets a clear name, a description a model can act on, typed inputs, and read-only or write annotations. Public data needs no sign-in; anything tied to a user's account goes through OAuth 2.0, so the assistant acts as that user and nothing more.
Listing in AI Directories
People find connectors in directories, so a server nobody lists is a server nobody uses. I prepare and submit to Anthropic's Connectors Directory for Claude, OpenAI's plugin directory for ChatGPT and Codex, and the official MCP Registry, plus community directories such as Smithery, Glama and PulseMCP. Each has its own checklist (tool annotations, privacy policy, OAuth setup, domain verification, reviewer test accounts), and I work through it and fix whatever review flags.
WebMCP for In-Browser Agents
WebMCP is a proposed web standard that lets a page register its own actions (search, filter, add to cart, submit a form) as tools an AI agent in the browser can call directly, instead of guessing at buttons. Chrome is testing it as an origin trial, so I add it as progressive enhancement: agents that support it get clean tools, and every other visitor gets the same site as before.
AI Crawler Access
Plenty of sites block AI crawlers without knowing it, through a CDN's managed robots.txt, a firewall rule or a bot-fight setting. I check what GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot and Google-Extended actually get from your site, then set a deliberate policy: let in the search and citation bots you want, block the ones you don't.
llms.txt and Readable Pages
An llms.txt that points models to your key pages, generated from your own content so it never goes stale, and pages that put the answer in server-rendered HTML rather than behind JavaScript. A straight note: llms.txt is cheap to add, but no major search engine has said it uses the file to rank or cite pages, so it's one small step, not the strategy.
Structured Data and AI Search Visibility
Schema.org JSON-LD for your organization, services, products, FAQs and articles, so AI answers state facts about you correctly. Then the GEO work: find the questions people ask assistants in your market, check whether you're cited, and write pages that give a direct, quotable answer.
Agent-Readable API
If you already have an API, an OpenAPI description that agents and tool builders can read, consistent error messages a model can recover from, and rate limits that don't treat every automated request as an attack. Forms get proper labels and clear validation messages, which helps browser agents and screen readers alike.
Live examples
MCP servers I run in production
A catalog of real objects photographed on white, built to be used by AI agents. The public MCP server needs no account: agents search the catalog, fetch an object with its license and credit line, or request one that's missing. Every tool carries read-only or write annotations, and the server is published in the official MCP Registry under the domain-verified name com.obbjects/obbjects.
Seoul real-estate platform with an MCP server running alongside the website. Assistants can search apartment and officetel listings, open a listing's full price breakdown, find verified experts and send an inquiry. Signed-in users connect through OAuth to save listings and set price alerts on their own account.
The site you're on. Its llms.txt and llms-full.txt are generated from the same data as the pages, and its robots.txt overrides Cloudflare's managed AI-crawler block for search and citation bots (GPTBot, OAI-SearchBot, ClaudeBot, PerplexityBot) while still blocking Bytespider.
How it works
From audit to listed
- 01
Audit
I check what AI crawlers and assistants can see today: robots.txt and CDN bot rules, the rendered HTML, structured data, and whether assistants mention or cite you for the questions that matter in your market. You get a short written list, ranked by impact.
- 02
Tool design
We pick the handful of actions an assistant should be able to take on your product and write them as tools: names, descriptions, inputs, what's read-only, what needs sign-in. This step decides whether models call your tools correctly, so it's done before any code.
- 03
Build
The MCP server on your existing stack, OAuth where it's needed, WebMCP on the pages where it helps, plus the crawler, llms.txt and schema fixes from the audit. Tested in Claude, ChatGPT and MCP Inspector with real prompts, not just unit tests.
- 04
Listing and follow-up
Submissions to the directories that fit, under your company's accounts, and fixes for anything reviewers flag. After launch I watch tool-call errors and AI citations, and rewrite tool descriptions where models misuse them.
FAQ
Common questions
- What is an MCP server, and why would my website need one?
- MCP (Model Context Protocol) is the open standard that Claude, ChatGPT, Cursor, VS Code and most other AI assistants use to connect to outside tools. An MCP server for your site is a small service that exposes your product's actions (search listings, check prices, book, look up an order) as tools the assistant can call. Without one, an assistant can only read your public pages, if it can reach them at all. With one, a user can ask for a two-bedroom apartment under their budget and the assistant runs your search directly.
- Do you have to rebuild my website?
- No. The MCP server sits next to your existing site and calls the same database or API it already uses. It works with Next.js, WordPress, Shopify, Rails, Django or a custom backend; what matters is that your data and actions can be reached from a server. Changes to the site itself are limited to crawler rules, structured data, llms.txt and, if you want it, WebMCP.
- Can you get my MCP server listed in Claude's and ChatGPT's directories?
- I prepare the submission and take it through review, but approval is each platform's call, so nobody can honestly guarantee a listing. Anthropic's Connectors Directory takes remote MCP servers from any paid Claude plan through its developer portal; submissions are scanned automatically and listed as a Community connector by default, and some also get a human review. OpenAI's directory, shared by ChatGPT and Codex, takes MCP servers packaged as plugins and asks for a verified developer identity, domain verification and reviewer test cases. The official MCP Registry stores metadata only and lists a server once you prove you own its namespace. Submissions go under your company's accounts, because the listing should belong to you.
- What is WebMCP, and is it ready to use?
- WebMCP is a proposed web standard, still being incubated in a community group, that lets a web page declare its own tools to an AI agent running in the browser. Chrome is testing it as an origin trial, and the API is still moving (navigator.modelContext is giving way to document.modelContext). I add it as progressive enhancement behind feature detection, so it costs your regular visitors nothing and is easy to update as the spec settles.
- Does llms.txt actually help?
- It's cheap and harmless, and some AI tools do fetch it, but no major search engine has said it uses llms.txt to rank or cite pages. What moves AI search visibility is the same as for regular search: crawlers allowed in, answers present in the server-rendered HTML, accurate structured data, and pages that answer the question directly. I add llms.txt as part of that work, not instead of it.
- Is it safe to let AI assistants act on my site?
- It's as safe as the tools you expose. Read-only tools such as search and look-up carry little risk. Anything that changes data goes through OAuth, so it runs as a signed-in user with that user's permissions, is marked as a write action so the client can ask the user to confirm, and is validated and rate-limited on the server like any other request. Admin actions and raw database access are never exposed to an assistant.
- How is a project priced?
- Each project gets a fixed quote once I've seen your site and we've agreed the tool list. The main cost drivers are how many tools you need, whether they need user sign-in (OAuth adds work), how clean your existing API or database is, and how many directories you want to be listed in. Send your URL and what you want assistants to be able to do, and I'll reply within 48 hours with a scope and a quote.
Want AI assistants to use your product?
Send your URL and the actions you'd want an assistant to take. I'll reply within 48 hours with a proposed tool list and a quote.
Get in touch →