Skip to content

MCP SERVER DEVELOPMENT

MCP server development: custom MCP servers that let AI work in your CRM, ERP and data

We build a Model Context Protocol server your company owns: a short list of clearly described tools over your systems, behind OAuth 2.1 sign-in. Claude, ChatGPT or your own agent can then look up an account, check an order or draft a task, always with the permissions of the person asking and a log entry for every call.

  • MCP servers in production
  • OAuth 2.1 sign-in, per-user permissions
  • Building software since 2012

Free consultation

Get a proposal for your project

Tell us what you want to automate. An engineer will read it and reach out to set up a free consultation.

We need an email or a phone number to get back to you.

THE PROBLEM

Your team already uses AI assistants. They just can’t see your systems.

Your people ask Claude and ChatGPT business questions every day, but the answers live in your CRM, ERP and databases. Without a controlled connection, they fill the gap by hand or with shortcuts nobody signed off on.

Copy, paste, repeat

An account manager at a Houston plumbing supplier exports open orders to a spreadsheet, pastes them into a chat window and asks which ones are late. Tomorrow she’ll do it all over again.

One integration per tool

Operations wants ChatGPT, engineering prefers Claude and IT is piloting an in-house agent. Hooking each one up to the ERP separately triples the work and the attack surface.

A shared key with full access

Someone connected an assistant with an admin API token. Now every user can read every record, and the logs show the token, not the person behind the question.

HOW AN MCP SERVER WORKS

Follow one question from an AI assistant to your ERP and back

An MCP server sits between AI clients and your systems and decides what each request is allowed to do. In this illustrative example, an account manager asks her assistant about a customer’s open orders. The assistant finds the server’s tools, she signs in and approves it, the call gets checked against her own permissions, the ERP answers and both steps land in the audit log.

Illustrative example of one request through a company MCP server, for a fictional plumbing supply distributor in Houston. Dana, an account manager, asks her AI assistant which open orders Rivera Plumbing has and whether any are late. The assistant connects to the company’s remote MCP server and lists the tools it offers: search accounts and get open orders, which only read data, and create a follow-up task, which needs her confirmation. Dana signs in with her company account and approves the assistant, with read access to accounts and orders only. The assistant calls get open orders with the customer and the status as arguments. The server checks that the call acts with Dana’s own permissions: she is signed in, the orders scope was granted, the customer is in her territory and she is under the rate limit. The server then queries the ERP, gets three open orders, one of them past its ship date, and returns them to the assistant, which answers Dana and offers to create a follow-up task. The consent and the tool call are written to the audit log.

Illustrative example Fictional business, people and data. Not a real client, conversation or result.

REMOTE MCP SERVER

One remote MCP server between every AI client and your systems

A remote MCP server runs on the web, not on each employee’s laptop, so you deploy it once, control who signs in and update tools in one place. AI clients connect over HTTPS. The server holds the connections to your systems and never hands out their credentials.

  • Connected to Your MCP server: Claude (AI clients)
  • Connected to Your MCP server: ChatGPT (AI clients)
  • Connected to Your MCP server: Your own AI agent (AI clients)
  • Connected to Your MCP server: CRM (Business systems)
  • Connected to Your MCP server: ERP (Business systems)
  • Connected to Your MCP server: Help desk (Business systems)
  • Connected to Your MCP server: SQL database (Data)
  • Connected to Your MCP server: Internal APIs (Data)
  • Connected to Your MCP server: Sign-in (OAuth 2.1) (Identity)
Illustrative architecture. Which clients and systems connect depends on your accounts, plans and APIs, which we verify during scoping.
ENTERPRISE MCP

Enterprise MCP, built on the stack we already run in production

Our team builds and operates MCP servers in production with Laravel and its first-party MCP package (laravel/mcp), behind OAuth 2.1 authorization. That’s the pattern we bring to your project, not a framework we’re trying out for the first time.

Each AI client registers itself with the server (dynamic client registration), the person signs in with their own account and approves it, and from then on every tool call acts with that person’s permissions. Staff and customers get separate servers, so one kind of account can never reach the other’s tools. If you also need AI inside your own product, our AI integration services cover direct API work alongside the server.

  • OAuth 2.1 with dynamic client registration
  • Per-person consent and permissions
  • Separate servers per audience
  • Rate limits and a full audit log
CHATGPT MCP · MCP FOR CLAUDE · MCP AGENTS

One MCP server for Claude, ChatGPT and your own agents

The whole point of the protocol is that you build the server once and every compatible client can use it. In practice, each client has its own setup, plan requirements and admin controls, so we test with the ones your team actually uses.

01 / 04

MCP for Claude: your tools as a custom connector

Claude can connect to remote MCP servers as connectors, so an employee can ask about an account or an order without leaving the conversation. Availability depends on your Claude plan and what your workspace admins turn on. We confirm the current options for your account before we build.

  • Each person signs in with their own account
  • Admins decide which connectors are allowed
  • Tool descriptions written for how Claude chooses tools
02 / 04

ChatGPT MCP: the same server, a second client

ChatGPT also supports MCP-based connections, with options that vary by plan and workspace settings. Because the server follows the standard, adding ChatGPT doesn’t mean a second integration with your ERP. It means testing, client-specific settings and a sign-off from whoever manages your ChatGPT workspace.

  • No second integration with your systems
  • Same permissions and audit log
  • Workspace policies reviewed with your IT team
03 / 04

MCP agents: your in-house agent uses the same tools

An MCP agent is an AI agent that discovers and calls tools through MCP servers. If you’re building one, for a back-office workflow or inside your product, it can use the same server the assistants use, so tools, permissions and logs stay in one place. See how we design custom AI agents end to end.

  • Model-agnostic: works with major model APIs
  • Service accounts with their own narrow scopes
  • Human confirmation before consequential actions
04 / 04

Several clients, one place to govern them

Most companies end up with a mix: one team on Claude, another on ChatGPT, an internal agent in development. With one MCP server in the middle, you approve tools once, revoke a client in one place and read one audit log, whatever the next AI tool turns out to be.

  • Approve or revoke clients centrally
  • One audit log across clients
  • Add tools without touching each client
MCP SECURITY AND MCP OAUTH

MCP security: the controls built into every server we ship

An MCP server is a door into your systems, so security is most of the engineering. These controls come standard, and your security team reviews them before launch, not after.

OAuth 2.1 authorization

Clients register dynamically, people sign in and approve the client, and the server issues short-lived tokens. No shared admin keys.

The person’s own permissions

Every tool call runs as the signed-in person, so the AI sees exactly what that person could see in the system itself, and no more.

Separate servers per audience

Staff and customers connect to different servers, so a customer account can’t reach internal tools, even by mistake.

Narrow, described tools

“Get open orders for a customer,” not “run any query.” Each tool has a clear description and a strict input schema.

Confirmation before writes

Read tools first. Tools that change data ask for explicit confirmation, and the riskiest ones stay out of the server entirely.

Rate limits and audit log

Limits per person and per client, plus a log of who called which tool, with what arguments and what came back.

COMPARISON

Custom MCP server vs custom API integration vs platform actions

Choose a custom MCP server when several AI assistants or agents need governed access to the same systems. Choose a custom API integration when AI is one feature inside your own software. Platform-specific actions, like custom GPT actions, are the quickest route when one team works in a single AI platform.

  • Best for

    Custom MCP server (Recommended)
    Many AI clients, the same systems
    Custom API integration
    One AI feature inside your product
    Platform actions (e.g. custom GPT actions)
    One team on one AI platform
  • Works with Claude, ChatGPT and in-house agents

    Custom MCP server (Recommended)
    Yes
    Custom API integration
    No
    Platform actions (e.g. custom GPT actions)
    No
  • Each person acts with their own permissions

    Custom MCP server (Recommended)
    Yes
    Custom API integration
    Partly
    Platform actions (e.g. custom GPT actions)
    Partly
  • Tools self-described to the model

    Custom MCP server (Recommended)
    Yes
    Custom API integration
    Partly
    Platform actions (e.g. custom GPT actions)
    Yes
  • Your own audit log of every call

    Custom MCP server (Recommended)
    Yes
    Custom API integration
    Yes
    Platform actions (e.g. custom GPT actions)
    Partly
  • Works with in-house or legacy systems

    Custom MCP server (Recommended)
    Yes
    Custom API integration
    Yes
    Platform actions (e.g. custom GPT actions)
    Partly
  • Time to first version

    Custom MCP server (Recommended)
    Weeks
    Custom API integration
    Weeks
    Platform actions (e.g. custom GPT actions)
    Days
  • Main trade-off

    Custom MCP server (Recommended)
    A young standard; client support still varies
    Custom API integration
    Rebuilt for each new AI tool
    Platform actions (e.g. custom GPT actions)
    Tied to one vendor’s format and rules

General comparison. Many projects combine an MCP server for assistants with direct API calls for product features.

HONEST LIMITS

When an MCP server isn’t the right answer

MCP solves a specific problem: governed access for AI clients you don’t control. In these situations, we’ll recommend something else, or a smaller first step.

One feature, one app

If AI only needs to summarize records inside your own software, a direct API integration is simpler and you skip running a server.

No API underneath

An MCP server wraps existing access. If the system has no API, database access or export, that problem has to be solved first.

Untrusted content in the loop

A model that reads emails or web pages can be steered by hidden instructions (prompt injection). We keep write tools narrow and confirmed, but no design makes that risk zero.

A standard still moving

The protocol and each client’s support for it keep evolving. Budget for maintenance: spec updates, client changes and re-testing are part of owning a server.

CUSTOMIZATION

What you decide about your MCP server

Your policies shape every server. We write these decisions down during scoping and test each one before anyone connects.

  • Tools: the exact operations exposed, named and described in plain language
  • Audiences: which servers exist, such as staff only, partners or customers
  • Scopes: what each role may read and which actions need confirmation
  • Allowed clients: which AI assistants and agents may register
  • Data rules: fields masked or never returned, such as personal or financial details
  • Limits: rate limits per person, per client and per tool
  • Language: tool descriptions and responses in English, Spanish or both
  • Logging: what the audit log keeps, where and for how long
SYSTEMS AND MCP SERVER HOSTING

What your MCP server connects to, and where it runs

These are the categories we’re usually asked to expose. We confirm access for your specific accounts, plans and versions before we quote.

AI clients

  • Claude
  • ChatGPT
  • In-house agents on model APIs

Business systems

  • CRM
  • ERP
  • Help desk
  • Accounting

Data

  • PostgreSQL
  • MySQL
  • SQL Server
  • REST and GraphQL APIs

Identity and access

  • OAuth 2.1
  • Dynamic client registration
  • Your existing user accounts

Hosting

  • Your cloud account
  • Infrastructure we operate for you
HOW WE BUILD AN MCP SERVER

How we build an MCP server, from tool list to production

We start with a short list of read-only tools, prove they answer real questions, and only then add anything that writes.

Step 1 of 5

Tool design

We collect the questions your team asks AI today and turn them into a short list of narrow tools over specific systems.

Deliverable: a tool catalog with inputs, outputs, scopes and the systems behind each tool.

Step 2 of 5

Access model

We map roles to scopes, decide which audiences get their own server and agree on what gets masked or never returned.

Step 3 of 5

Read-only server

A remote server with OAuth 2.1 sign-in goes up in a staging environment, connected to test or read-only data.

Step 4 of 5

Test with real clients

Your team uses it from Claude, ChatGPT or your agent on real questions. We tune tool names and descriptions until the right tool gets picked.

We don’t talk about write tools until your team signs off on the read-only results.

Step 5 of 5

Launch and operate

Production deployment, rate limits, alerts and audit log, then protocol and client updates re-tested as they ship.

EXAMPLES

What companies expose through a custom MCP server first

Most first servers answer questions people already ask several times a day. Illustrative scope, not reported results.

  • B2B distribution

    Account and order status

    Sales reps ask their assistant about a customer’s open orders, balances and last contact without opening the ERP or the CRM.

  • E-commerce

    Order context for support

    Support agents pull an order, its shipment and past tickets into the conversation, then draft a reply they review.

  • Dental

    Open slots for the front desk

    Front-desk staff ask which hygiene slots are open this week, and the server returns availability without patient details.

  • Physical therapy

    Plan-of-care schedule checks

    Clinic coordinators check which patients still need sessions booked in their plan, with only the fields their role allows.

  • Healthcare

    Intake status by role

    Office staff ask which intake forms are still missing. We review the data rules for health information on every project.

  • Restaurants

    Reservations and covers

    Managers ask for tonight’s large parties and special requests across locations from the assistant they already use.

WHY APPYWEB

Why companies choose us to build their MCP server

We’re not new to the protocol. Our team builds and runs MCP servers in production, with OAuth 2.1 sign-in and per-user permissions, and the engineers who scope your server are the ones who build it.

Running in production

The patterns on this page come from servers we operate, not from a demo.

Read-only first

We prove the tools answer real questions before anything can write.

Documented

Tool catalog, data flows and runbooks your security team can review.

Austin + Spain

Headquartered in Austin, Texas, with engineers in Spain.

RELATED SERVICES

Related AI development services

FAQ

MCP server development questions

What does MCP server development include?

Designing the tools (names, descriptions, input schemas), connecting them to your systems, OAuth 2.1 authorization, permissions per role and per person, rate limits, an audit log, deployment as a remote server, testing with the AI clients your team uses, and documentation for your security review. After launch it also means maintenance as the protocol and the clients evolve.

How long does it take to build a custom MCP server?

It depends on how many systems and tools are involved, the state of their APIs and the approvals your security team needs. A read-only server over one well-documented system is much faster than write tools across several. You get a written plan with phases after the tool design step.

How does MCP OAuth work?

For remote servers, the MCP specification builds on OAuth 2.1. The AI client registers itself with the server (dynamic client registration), the person is sent to sign in with their own account and approves the client for specific scopes, and the client receives a token. Every tool call then carries that token, so the server acts with that person’s permissions and logs the call under their name.

Is MCP secure?

The protocol gives you structure; security depends on how the server is built. The main MCP security risks are over-broad tools, shared credentials, prompt injection from content the model reads, and unvetted third-party servers. We address them with OAuth 2.1, per-person permissions, separate servers per audience, narrow tools, confirmation before writes, rate limits and an audit log, and we’re upfront that prompt injection can be reduced but not ruled out.

What is a remote MCP server?

A remote MCP server runs on the web and clients connect to it over HTTPS, instead of a local server that runs as a program on each user’s computer. Remote servers suit companies: one deployment, central sign-in and permissions, one audit log, and updates that reach every user at once. Local servers still make sense for developer tools that work with files on one machine.

Where is the MCP server hosted?

Where your security team prefers: in your own cloud account, next to the systems it connects to, or on infrastructure we operate for you. Either way it sits behind HTTPS, keeps the credentials for your systems on the server side and is monitored for errors, latency and unusual usage.

Can we use the same MCP server with ChatGPT and Claude?

Yes, that’s the main reason to build one. Claude and ChatGPT both support MCP-based connections, and in-house agents can use the same server. Each client has its own setup, plan requirements and admin controls, so we test with the clients and plans your team actually uses and document the settings for each.

Can an MCP agent make changes in our systems, or only read data?

It can, if you expose write tools. We start read-only, then add narrow write tools such as creating a follow-up task or a draft, with explicit confirmation from the person before anything is saved. Destructive or high-risk operations usually stay out of the server entirely.

Do we still need our own API if we build an MCP server?

Yes. An MCP server usually sits on top of your existing APIs or database access and makes them usable by AI clients, with descriptions, schemas and permissions. Your API stays the foundation for your apps and integrations, and the MCP server is the controlled layer that faces AI. Our guide to the Model Context Protocol on the blog compares the two in more detail.

Tell us what your AI assistants should be able to reach

List the systems, the questions your team asks AI today and which assistants they use. We’ll get back to you with a first tool list, the access model we recommend and what a read-only first version would include.

  1. 01 We read what you sent before we talk.
  2. 02 A free call to map the workflows worth automating.
  3. 03 A written proposal with scope and a fixed price.

We need an email or a phone number to get back to you.