AI services

MCP Servers & AI Integrations

Put your product or data in front of ChatGPT, Claude and other AI hosts through the Model Context Protocol. Secured with OAuth 2.1 and tested against each host's rules.

At a glance

MCP servers for search, data and actions in your product
OAuth 2.1 with Entra, APIM or your identity provider
Conformance with the rules of each AI host
Load, security and evaluation testing before launch

An MCP server is the program that lists a product's tools, such as search, read a record or create an order, so an AI agent can call them on a person's behalf. Momentaj, a Toronto AI engineering firm, designs, builds and tests MCP servers that let AI agents search and act on a company's product or data. Your website and your product have a new kind of visitor: AI agents. The Model Context Protocol (MCP) is how ChatGPT, Claude, VS Code, Gemini CLI and a growing list of AI hosts let an agent search, read and act on an outside system. An MCP server decides what an agent can see and do with yours, and who controls that.

Who it's for

SaaS companies whose customers already ask ChatGPT or Claude questions the product could answer
Marketplaces, booking and commerce businesses that want agents to search and transact, not scrape
Enterprises exposing internal data and tools to their own AI assistants, safely

MCP server, API or plugin?

An MCP server serves a different caller from an API or a plugin, and it usually sits on top of the API you already have.

Option	Who decides what to call	Reach
API, for programs a developer writes	The developer, in code	Every developer who integrates
Plugin in one host's own format	Depends on the host	One host; each further host is another build
MCP server, for any AI agent	The model, on each request, from the tool descriptions	Every host that speaks MCP, each with its own rules

Most of the work is tool design over that API: three to six tools for a first server, not one per endpoint. A plugin built only in one host's own format is rarely the right first build for AI agents. A plugin can also carry an MCP server: ChatGPT and Codex plugins can include one, and many of the website tools in our founder's survey of MCP servers for websites run one. When you need something only one host offers, build it around an MCP server you can reuse.

What we build

Remote MCP servers that expose search, discovery, records and actions from your existing APIs, with the tool design that makes an agent use them well.
Authentication and authorization. OAuth 2.1 with PKCE, Entra External ID or your identity provider, API Management in front, scoped tokens, and per-user permissions carried through to every tool call.
MCP Apps and rich results. Tool results that render as cards, lists and forms inside the host, where the host supports it.
Host compliance. Tool names, annotations, descriptions, timeouts and payload sizes that satisfy the different rulebooks of the MCP specification, Claude, ChatGPT, VS Code, Cursor, Kiro and Gemini CLI, and the connector-directory submissions that follow.
Testing and operations. Inspection, conformance checks, evaluation sets, load tests, OpenTelemetry tracing and cost dashboards.

How it's done

We start from the questions your users ask an agent, not from your API surface. A good MCP server has few tools, each named and described so a model picks the right one, each returning compact, well-typed results. We prototype against a real host within the first two weeks, then harden: auth, rate limits, observability, abuse cases.

Our founder leads the architecture of a global hospitality brand's guest-facing MCP platform, which exposes hotel search, discovery and booking through ChatGPT and Claude, and publishes comparisons of MCP server tooling, testing tools and host requirements that we use in our own work.

How an MCP server is secured

Securing an MCP server means controlling who the agent acts for, what it may do for that person, and what it may believe from the text it reads.

OAuth 2.1 with PKCE. A remote server acts for one person, never through a shared service account, and accepts only tokens issued for itself.
The person's permissions in every tool call. Scoped tokens get the agent in; the check that matters runs inside each tool, so the agent can never do more than the person it acts for.
Separate read and write tools. Reading can be granted without writing, and search and fetch tools come before actions.
Honest hints. Every tool has a title and read-only, destructive and open-world hints that match what it actually does.
Approval before anything irreversible. The agent proposes and a person approves, through an approval gate.
Injection through tool results. A record the server returns may carry instructions someone else wrote, and the model cannot reliably tell them from data. Telling the model to ignore them helps but does not hold; server-side grounding, permission checks outside the model and small results do. Run prompt injection tests before every release and again when the model changes.

Proof

Our notes on this site set out what we apply in a build: Should your product have an MCP server? A decision guide, What each AI host requires of an MCP server, Testing an MCP server: what exists and what is still missing and Your website has a new visitor: AI agents, all collected in our notes on MCP servers. The full research behind them, including a comparison of 44 MCP tools for websites, the rules eight AI hosts apply to an MCP server, and 31 MCP testing tools compared, is at pournasserian.com/writing.

Engagement shape

A two-week design sprint that produces a tool design, an auth design and a working prototype against one host, then a production build. We work remotely with clients across Canada, the United States, the Middle East and beyond, and in person across the Greater Toronto Area. See how an engagement works.

Notes on this

Frequently asked questions

What is an MCP server, and how is it different from an API?

An MCP server is the program that lists a product's tools, such as search, read a record or create an order, so an AI agent can call them on a person's behalf. It does not replace your API: it usually sits over the API you already have, and most of the work is tool design, a few tools each named and described so a model picks the right one and each returning compact results.

Should our product have an MCP server?

If your customers already use ChatGPT or Claude to ask questions your product could answer, or if agents are scraping your site to do it badly, yes. If nobody is asking yet, a short assessment tells you whether to build now or wait.

Which AI hosts should an MCP server support?

For most products, ChatGPT and Claude first, because of reach and because their connector directories bring users. Developer products add VS Code, Cursor and the CLIs. Each host applies its own rules, which we design for from the start.

How does an MCP server handle authentication?

OAuth 2.1 with PKCE is the standard the hosts expect. We put an identity provider such as Entra External ID behind it, an API gateway in front, and carry the user's permissions into every tool call, so an agent can never do more than the person it acts for.

How do you test an MCP server?

Protocol conformance, each host's rules, evaluation sets that check the model picks the right tool with the right arguments, load tests, and security tests for prompt injection through tool results.

How much does an MCP server cost, and how long does it take?

An MCP server is quoted in two steps, a design sprint and then a production build, because the cost depends on your API and the hosts you target. The design sprint of about two weeks produces the tool design, the authentication design and a working prototype against one host; the production build is then quoted at a fixed price. If your API is clean, the whole build takes weeks rather than months; if the API needs fixing first, that comes before the server.

Talk to the people who would build it

Tell us what you are trying to build. We will come back with an honest take and a plan.

Book an AI consultation