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.
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.
Free consultation
Tell us what you want to automate. An engineer will read it and reach out to set up a free consultation.
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.
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.
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.
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.
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 Fictional business, people and data. Not a real client, conversation or result.
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.
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.
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.
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.
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.
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.
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.
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.
Clients register dynamically, people sign in and approve the client, and the server issues short-lived tokens. No shared admin keys.
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.
Staff and customers connect to different servers, so a customer account can’t reach internal tools, even by mistake.
“Get open orders for a customer,” not “run any query.” Each tool has a clear description and a strict input schema.
Read tools first. Tools that change data ask for explicit confirmation, and the riskiest ones stay out of the server entirely.
Limits per person and per client, plus a log of who called which tool, with what arguments and what came back.
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.
General comparison. Many projects combine an MCP server for assistants with direct API calls for product features.
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.
If AI only needs to summarize records inside your own software, a direct API integration is simpler and you skip running a server.
An MCP server wraps existing access. If the system has no API, database access or export, that problem has to be solved first.
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.
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.
Your policies shape every server. We write these decisions down during scoping and test each one before anyone connects.
These are the categories we’re usually asked to expose. We confirm access for your specific accounts, plans and versions before we quote.
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
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
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
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
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
Production deployment, rate limits, alerts and audit log, then protocol and client updates re-tested as they ship.
Most first servers answer questions people already ask several times a day. Illustrative scope, not reported results.
Sales reps ask their assistant about a customer’s open orders, balances and last contact without opening the ERP or the CRM.
Support agents pull an order, its shipment and past tickets into the conversation, then draft a reply they review.
Front-desk staff ask which hygiene slots are open this week, and the server returns availability without patient details.
Clinic coordinators check which patients still need sessions booked in their plan, with only the fields their role allows.
Office staff ask which intake forms are still missing. We review the data rules for health information on every project.
Managers ask for tonight’s large parties and special requests across locations from the assistant they already use.
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.
The patterns on this page come from servers we operate, not from a demo.
We prove the tools answer real questions before anything can write.
Tool catalog, data flows and runbooks your security team can review.
Headquartered in Austin, Texas, with engineers in Spain.
Custom AI software, LLM and RAG systems, AI integrations and AI SaaS, built by in-house engineers.
Learn moreAI connected to your CRM, ERP, help desk and product through APIs, webhooks and MCP servers.
Learn moreEnterprise RAG, RAG chatbots and LLM applications that answer from your documents and cite the source.
Learn moreVoice, receptionist, sales, support and chat agents built around your processes and tools.
Learn moreDesigning 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.
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.
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.
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.
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 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.
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.
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.
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.
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.
Free AI audit
An engineer will review your answers and reach out to walk you through where AI can help most.
Pick the area you want us to look at. We’ll review how it runs today and tell you what to automate first, and what to leave alone.
Free, with no strings attached. An engineer reads your answers before reaching out.
Free consultation
Tell us what you want to automate. An engineer will read it and reach out to set up a free consultation.
Thanks, we got your request
Someone on our team will review it and reach out to schedule your free consultation.