// engineering · ai·№ 04 of 07

// engineering · new · № 04 of 07

Give your AI real access to your business data.

Pixelworx builds custom MCP servers that give AI assistants like Claude structured, authenticated access to your databases, APIs, and internal tools — so it works with your real data instead of whatever you paste into a chat window.

Discuss Your MCP Server → What we build ↓

mcp.stream live

13:42 tools.crm tool → query.contacts · auth ok mcp.4
13:41 resource.docs resource → policy.md · 2.4kb mcp.4
13:39 transport sse → /messages handshake mcp.4
13:36 audit.log log → tool.call · 14 today mcp.4
13:30 tools.invoice tool → extract.line-items mcp.4
13:24 studio note → mcp v0.4 spec compliant —

$ awaiting next tool call spec v0.4 compliant

// mcp explainer

What is an MCP server — and why does it matter for your business?

MCP stands for Model Context Protocol — an open standard created by Anthropic that defines how AI assistants communicate with external systems. Without one, using AI means copying data into a chat window every session. With a custom MCP server, your AI assistant queries your database directly, pulls CRM records on demand, searches your knowledge base, triggers workflows, or runs reports — all within defined, authenticated, auditable boundaries. MCP servers expose tools (actions the AI can invoke) and resources (data the AI can read). The model reasons about which tools to use based on context — rather than following hardcoded logic.

Talk to us about MCP →

Without an MCP server

Copy-paste data into every chat session
AI reasons over incomplete, stale context
No audit trail — no record of what the AI accessed
Generic answers from training data, not your actual records
Every session starts from zero

With a custom MCP server

AI queries your database and APIs in real time
Structured access with authentication and logging
Full audit trail of every tool call the AI makes
Answers grounded in your actual business data
AI persists knowledge of your business across sessions

// what we build

What a custom MCP server includes.

A production MCP server is more than a handful of tool definitions. We build the full stack — from connector logic to deployment to monitoring.

06 capabilities

Tool and resource definitions

The structured interface the AI uses to understand what's available and how to use it. We design tool definitions that are clear to the model, safe by default, and specific enough to prevent misuse or ambiguous calls. Good tool design is the difference between an AI that works reliably and one that guesses.

Tool schemas and descriptions Input validation Response formatting Error handling

Database and API connectors

The actual integration logic — queries, API calls, data transformations, and response formatting. Built to handle real data volumes, edge cases, and partial failures gracefully. Whether the source is MySQL, PostgreSQL, a REST API, or a third-party service, we build connectors that are reliable under production load.

MySQL / PostgreSQL REST API connectors Data transformation Graceful failure handling

Authentication and access control

MCP servers sit in front of your data. We implement proper auth — API keys, OAuth, scoped permissions — so only authorized AI sessions can access what they're permitted to access. Access is configurable per tool and per user context.

API key authentication OAuth support Scoped permissions Session-level access control

Audit logging

Every tool call the AI makes is logged — what was requested, what parameters were passed, what was returned, and when. Essential for debugging, compliance, understanding AI behavior, and catching unexpected usage patterns.

Request and response logging Parameter capture Timestamp and user tracking Queryable audit trail

Transport and deployment

Stdio transport for Claude Desktop and local tooling (Claude Code, Cowork). HTTP with SSE transport for networked or cloud-hosted deployments. We handle the packaging, deployment, environment configuration, and process management for whichever mode fits your use case.

Stdio (local / desktop) HTTP + SSE (networked) Cloud deployment Process management

Documentation and handoff

Clear documentation of every tool, resource, and connector — what it does, what it expects, what it returns, and what happens when it fails. So your team can maintain, extend, or hand off the server without reverse-engineering our work.

Tool reference docs Deployment runbook Extension guide Troubleshooting reference

// use cases

Real-world MCP server use cases.

MCP servers work wherever AI needs to act on real business data rather than generic training knowledge. These are the patterns we see most often.

08 patterns

CRM assistant — Claude queries your HubSpot or Salesforce records and drafts follow-up emails based on actual deal history

Internal knowledge base — Claude answers questions from your actual documentation, policies, and SOPs — not hallucinated approximations

Inventory and product lookup — Claude checks live stock levels, pricing, and availability through your ERP or database

Report generation — Claude queries your production database and produces formatted reports on demand, without manual export

Support triage — Claude reads existing ticket history and customer records before suggesting a resolution path

Developer tooling — Claude Code accesses your codebase context, architecture docs, and runbooks through a custom MCP server

Supplier data access — Claude pulls vendor pricing, lead times, and compliance data from your supplier management system

Project management — Claude checks task status, resource assignments, and blockers across your project data without leaving your workflow

// when to build custom

When does your business actually need a custom MCP server?

Off-the-shelf MCP connectors cover generic workflows. A custom MCP server is the right call when any of these apply.

06 signals

Your data lives inside systems the AI cannot see

If the information you need the AI to work with lives in your CRM, ERP, database, or knowledge base — a custom MCP server is how you bridge that gap. Without it, the AI only knows what you paste into a chat window.

You need scoped, auditable AI access

Compliance, governance, or internal policy may require a clear record of what the AI accessed, when, and with what parameters. A custom MCP server gives you audit logging, scoped permissions, and access control that generic connectors cannot provide.

Your business logic is too specific for generic connectors

Prebuilt connectors assume standard workflows. If your data model, pricing rules, or CRM configuration doesn't match the generic template, custom tool definitions let you expose data the way your business actually works.

You want one AI interface across multiple systems

A single MCP server can expose tools that reach your CRM, your database, your docs, and your APIs — so the AI has the full picture without you managing multiple connectors or feeding it context from each system separately.

You're building an AI-native product

If MCP is part of your product architecture — not just an internal tool — you need a server engineered for production: multi-tenant access control, rate limiting, and a deployment posture your customers can depend on.

You need a production deployment, not a local prototype

Many MCP servers start as Claude Desktop experiments. A production server — hosted, secured, monitored, and maintainable by your team — requires a different level of engineering. That's what Pixelworx builds.

// how it works

How MCP server development works.

We prototype fast and engineer properly — in that order.

04 phases

Define the tools

We map what the AI needs to access — which data sources, which actions, which queries. Tool design happens before any code, because the quality of tool definitions determines how well the AI uses them.

Build and validate

A working MCP server with the core tools is built and tested against your actual data. We validate that the AI reasons correctly — that it calls the right tools, interprets responses accurately, and stays within scope.

Add auth, logging, hardening

Once the tools work, we add the production layer — authentication, access control, audit logging, error handling, and rate limiting. This is where an MCP server becomes trustworthy enough to deploy.

Deploy and document

We deploy with the appropriate transport (stdio or HTTP/SSE), configure the environment, verify the deployment, and hand off clear documentation — tool reference, deployment runbook, and extension guide.

// frequently asked

Frequently asked questions about MCP server development.

12 questions

What is an MCP server? +

MCP stands for Model Context Protocol — an open standard created by Anthropic that defines how AI assistants communicate with external systems. An MCP server is the software layer that sits between an AI model like Claude and your business data, APIs, and tools. It exposes structured, authenticated access points — called tools and resources — that the AI can call to retrieve data, trigger actions, or interact with your systems in a controlled, auditable way.

Why would my business need a custom MCP server? +

If you want an AI assistant that actually knows your business — your CRM records, your inventory, your project data, your internal knowledge base — rather than one you have to manually feed context to every session, you need an MCP server. It's the difference between an AI that works with generic information and one that works with your specific data in real time.

What can an MCP server connect to? +

An MCP server can connect to any system your business uses — relational databases (MySQL, PostgreSQL), REST APIs, internal tools, CRMs like HubSpot or Salesforce, file systems, third-party services, or custom business logic. Each connection is exposed as a named tool or resource that the AI can call. Pixelworx designs the tool definitions, implements the connectors, handles authentication, and builds in logging and access control.

How is an MCP server different from a regular API integration? +

A regular API integration connects two software systems. An MCP server connects an AI model to your systems in a way the model can reason about and act on. The tool definitions tell the AI what's available, what parameters to pass, and what the response means — so the AI can decide which tools to call based on context, rather than being hardcoded to a fixed workflow.

Does Pixelworx build MCP servers for Claude specifically, or other AI models too? +

MCP is an open standard created by Anthropic and primarily supported by Claude, but the protocol is open and adoption is growing. Pixelworx builds MCP servers designed for Claude-based deployments — including Claude in Cowork, Claude Code, and custom Claude integrations — and can advise on compatibility with other AI tooling as the ecosystem matures.

What transport does Pixelworx use for MCP servers — stdio or HTTP? +

Both, depending on the deployment context. Stdio transport is used for Claude Desktop and local tooling (Claude Code, Cowork). HTTP with SSE transport is used for networked, multi-user, or cloud-hosted deployments. We select and configure the appropriate transport for your use case and can support both in the same server if needed.

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

A focused MCP server with a well-defined set of tools and a clean data source can often be built and deployed in days. More complex servers — with multiple data sources, custom auth flows, nuanced tool definitions, and monitoring requirements — take longer. We prototype fast: get a working version in front of you quickly, validate it does what you need, then engineer for production.

How much does a custom MCP server cost? +

Cost depends on the number of tools and data sources, the complexity of the connectors, the authentication requirements, audit logging needs, and the deployment posture. A focused server with one or two well-defined data sources is a contained project. A multi-source server with custom auth flows, monitoring, and documented handoff is more involved. We scope honestly before any commitment — tell us what systems you want the AI to work with and we'll explain what's realistic.

Can I run multiple MCP servers, and can they work together? +

Yes. Claude can connect to multiple MCP servers simultaneously, with each server responsible for its own domain — one for your CRM, one for your database, one for your internal docs. The model reasons across all connected servers and decides which tools to call based on the conversation context. There's no extra configuration required on your end; the client handles the multi-server connection.

What is the difference between MCP and function calling? +

Function calling is a model-specific feature baked into the OpenAI and Anthropic APIs — you define functions in the API request, and the model can call them. MCP is a standardized, transport-level protocol: any compliant AI client can connect to any compliant MCP server, regardless of the underlying model. MCP also supports resources (readable data objects), prompts (reusable message starters), and structured transport options that go beyond what function calling covers. For production, multi-user, or multi-client deployments, MCP is the better architecture.

Should I build a custom MCP server or use an off-the-shelf connector? +

Off-the-shelf connectors are the right choice when your workflow matches what they do and your data doesn't require custom auth or access control. Build custom when you have proprietary data in a non-standard schema, when you need audit logging that meets your compliance requirements, when your business logic is too specific for a generic tool definition, or when you need to expose multiple systems through a single controlled interface. We'll tell you plainly which path makes more sense for your situation.

Can Pixelworx add MCP capabilities to an existing Laravel application? +

Yes — this is a common engagement pattern. Laravel has first-class MCP support, and we can expose your existing Eloquent models, service classes, and APIs as MCP tools without a full rewrite. The MCP server can live inside your existing application or alongside it as a separate service. Either way, the integration uses your existing auth system and database — no new infrastructure required to get started.

// pairs well with

Related services.

MCP servers are one layer of a broader AI strategy. These are the services they most often connect to.

03 disciplines

// engineering · ai№ 03

AI development

The full scope of AI services — LLM integrations, RAG systems, AI automation pipelines, and AI-enhanced software beyond MCP servers.

Read discipline

// engineering№ 05

Third-party integrations

Connecting your CRM, ERPs, and APIs — often the same systems an MCP server exposes to AI. Clean integration work that makes data AI-accessible.

Read discipline

// engineering№ 01

Software development

The Laravel application layer MCP servers plug into — custom portals, SaaS platforms, and web applications built for production.

Read discipline

// further reading

MCP server articles from Pixelworx.

MCP server development from two angles — the business case and the engineering detail.

03 articles

business owners

What Is an MCP Server? A Guide for Business Owners

A plain-language explanation of what MCP servers are, why they matter, and how to evaluate whether your business needs one.

Read article →

saas founders & developers

Building MCP Servers for Your SaaS

Architecture decisions, tool definition patterns, and the engineering considerations for shipping an MCP server as part of a SaaS product.

Read article →

developers

One Action, Two Callers: The Laravel AI Action MCP Bridge

How to expose an existing Laravel AI Action as an MCP tool without duplicating code — using the pixelworxio/laravel-ai-action bridge.

Read article →

// ready when you are

Ready to give your AI real access to your business?

Tell us what systems you want the AI to work with and what you need it to do — we'll scope the MCP server, explain what's realistic, and get a working version in front of you fast.

Start a Conversation → [email protected]