Local MCP Creation: Build Custom AI Tools for Your Business in One Afternoon

What if AI could actually do your team's repetitive work — create CRM records, send follow-ups, coordinate between systems — not just answer questions about it? That is exactly what you can build now using the Model Context Protocol (MCP).

The breakthrough: you do not need developers. Using AI coding assistants like Cursor, Windsurf, or Claude Code, you describe what you want in plain English, and the AI builds it for you. We built a complete HubSpot AI assistant — managing companies, contacts, deals, and our entire sales workflow — in one afternoon.

This article shows you how.

____________________________

How to Build Your MCP Server: Step-by-Step

Step 1: Set up your AI coding assistant (15 minutes)

Download and install one of these:

• **Cursor** (cursor.sh) - What we used, free trial available

• **Windsurf** (codeium.com/windsurf) - Free option

• **Claude Code** - Part of Claude Pro subscription

Open the tool and create a new folder for your project. That is it. No other setup needed yet.

Step 2: Get your API credentials (10-20 minutes)

You need access to the system you want to automate. For HubSpot (our example):

• Log into HubSpot → Settings → Integrations → Private Apps

• Create a new private app

• Enable permissions for the objects you need (companies, contacts, deals)

• Copy the access token

Save this token somewhere secure. You will give it to the AI assistant later, and it will set up secure storage.

For other systems (Salesforce, Monday, Jira, etc.), the process is similar: find their API or developer settings, create credentials, copy the token or API key.

Step 3: Start the conversation with your AI assistant (30 minutes)

Open your AI coding assistant and start with this prompt:

"I want to build an MCP server that connects to [YOUR SYSTEM]. This server will let AI assistants perform actions in [SYSTEM] by using tools. Here is what I need it to do:

[Describe your workflow in plain English - see example below]

Please build this MCP server for me. Ask me any clarifying questions you need."

Our actual prompt for HubSpot:

I want to build an MCP server for HubSpot CRM. Here is our sales prospecting workflow:

When we find a potential customer:

1. Check if the company already exists in HubSpot (search by name)

2. If not, create a new company record with details we provide

3. Check if the contact person exists

4. If not, create a contact record

5. Create a deal with this format: '[Company Name] - Product Name'

6. Set default deal value to £15,000

7. Set close date to 3 months from today

8. Always link the contact to the deal (we need their email for outreach)

9. Add notes to the deal with source information

Also need to:

- Search for existing deals

- Update deal stages as they progress

- Look up our team members to assign deals

- Create notes on deals, companies, or contacts

Build this with proper error handling and make it work with Claude Desktop or Cursor as the AI client."

The AI will start asking questions: What properties do you want to capture? What deal stages exist? Any validation rules? Answer them based on how your team actually works.

Step 4: Let the AI build it (AI does this automatically - 10 minutes)

The AI will:

• Create the project structure

• Write the MCP server code

• Add authentication handling

• Build all the tools you described

• Set up error handling

• Create configuration files

You will see files being created. You do not need to understand them. The AI is doing all the work.

Step 5: Configure your credentials (5 minutes)

The AI will create a file for your API credentials (usually called `.env` or similar). It will show you where to paste your API token from Step 2.

Open that file and paste your token where indicated. Save it.

Step 6: Test each capability (1-2 hours)

This is the important part. Open the AI assistant (like Claude Desktop) that will use your MCP server, and start testing:

• Test searching:

You: "Search for companies called Acme in HubSpot"

AI: [Uses your tool, returns results]

• Test creating:

You: "Create a company called Beta Industries"

AI: [Creates it, returns confirmation]

• Test your workflow:

You: "Create a deal for Gamma Corp for our AI Helper product"

AI: [Executes your multi-step workflow]

When something does not work right, tell your AI coding assistant (Cursor/Windsurf/Claude Code): "When creating a deal, it should also [describe what's missing]." The AI updates the code immediately.

Common issues we fixed during testing:

• "Make sure it checks for existing companies before creating duplicates"

• "Add the job posting URL to the deal description"

• "Default currency should be GBP not USD"

• "Include the contact's job title when creating them"

Each fix took 30 seconds: describe the problem, AI fixes it, test again.

Step 7: Refine the instructions that guide AI behavior (30 minutes)

This is where the magic happens. The AI has created descriptions for each tool. These descriptions are instructions that guide how AI uses them.

Tell your AI coding assistant: "Show me the tool descriptions." You will see plain English instructions like:

"Search for companies in HubSpot by name or domain. Use this before creating a new company to avoid duplicates."

Refine these by saying: "Change that to say 'ALWAYS use this before creating a new company...'"

The stronger and more specific your instructions, the better AI follows your process. Examples:

• "ALWAYS search before creating"

• "Use exactly this format: [Company Name] - Product Name"

• "Default deal value for AI Helper is £15,000"

• "Close date is always 3 months from today"

• "Never create a deal without creating and linking a contact"

These are your business rules, written in sentences anyone can read and update.

Step 8: Deploy it (30 minutes - AI helps with this too)

Tell your AI coding assistant: "Help me deploy this so it runs reliably."

The AI will set it up to run either:

• Locally on your machine (simplest, good for testing and small teams)

• On a cloud server (for team-wide access)

For local deployment, the AI will show you how to start the server and connect it to Claude Desktop or your AI assistant.

For cloud deployment, the AI can create Docker configurations and deployment scripts. We deployed ours to Azure, but AWS or any other cloud works the same way.

Step 9: Use it in your daily work

Now when you talk to Claude (or your AI assistant), you can say:

• "Find all deals in the presentation stage"

• "Create a deal for this company I just found"

• "Search for existing contacts at Acme Corp"

• "Update this deal to the next stage"

The AI uses your custom tools automatically, following your business rules exactly.

Step 10: Iterate as your process evolves (ongoing)

When your workflow changes, talk to your AI coding assistant:

• "Add a new field to capture the lead source"

• "Change the default deal value to £20,000"

• "Add a new tool to export deals to CSV"

Each change is a conversation. The AI updates everything. You test it. Done.

Total timeline:

• Initial build: 3-4 hours

• Refinement over first week: 2-3 hours

• Ongoing updates: 15 minutes each

Compared to traditional development: 2-3 weeks for initial build, days for each change.

____________________________

Why Build Your Own Instead of Using Generic Tools

Your data stays under your control.

Your customer data and API credentials never leave your infrastructure. For regulated industries or any business serious about data privacy, this is not optional.

Generic tools do generic work. Custom tools do your work.

A standard connector creates deals. Our AI knows to use our naming format, set our default values, create the contact for outreach, and capture source information—because that is how our team works. This is your proprietary workflow, not someone else's template.

Business rules in plain English.

When your instructions say "ALWAYS search before creating," AI follows that every time. When your process changes, update the instruction. The AI adapts immediately. No developer backlog.

One instruction becomes a complete workflow.

Say: "Create a deal for Acme Corp from this job posting."

AI executes:

1. Reads and extracts company details from the posting

2. Searches HubSpot for existing company

3. Creates company if new

4. Searches for contact

5. Creates contact if needed

6. Creates deal with full context

7. Links everything together

8. Adds note with research details

Seven operations, coordinated automatically, from one sentence. This is what having an AI team member looks like.

The transformation happens when you connect AI to multiple systems.

Here are three scenarios:

Scenario 1: Job Boards + CRM (Sales Prospecting).

• Problem: Hours spent searching job boards, researching companies, manually entering data into CRM.

• Solution: One AI tool connecting job boards and your CRM.

• Say: "Find London companies hiring Dynamics 365 developers, create deals for the top 5."

AI automatically: searches jobs, extracts company details, checks CRM for duplicates, creates companies/contacts/deals with full context from job descriptions.

• Result: 2-3 hours weekly → 30 seconds.

Scenario 2: Email + CRM (Automated Follow-up).

• Problem: Deals stall because follow-ups get forgotten.

• Solution: One tool connecting CRM and email.

• Say: "Send follow-ups to all presentation stage deals not contacted in 2 weeks, then advance them."

AI automatically: finds matching deals, drafts personalized emails using deal context, sends through your email system, logs activity in CRM, advances pipeline stage.

• Result: Consistent, timely outreach. Zero forgotten deals.

Scenario 3: Project Management + CRM (Sales-to-Delivery Handoff).

• Problem: 30-60 minutes per closed deal to manually set up projects, copy information, create tasks.

• Solution: One tool connecting CRM and project management.

• Say: "Set up delivery projects for all deals closed this week."

AI automatically: finds closed deals, creates projects with tasks from your template, assigns team members, sets deadlines, links back to CRM.

• Result: Instant handoff with perfect accuracy.

Building these: Tell your AI assistant (Cursor/Windsurf/Claude Code): "Add job board capabilities to my HubSpot tool. When we find good postings, create deals automatically." The AI extends your existing tool. One conversation, two systems working together.

____________________________

Your Competitive Advantage: Speed + Control

The businesses dominating their markets in the next five years will not have the biggest AI budgets. They will be the ones turning business processes into AI-executable workflows faster than anyone else.

When you go from "we need AI to do X" to "AI is doing X in production" in one afternoon, you move faster than competitors waiting for engineering sprints.

When your AI tools encode YOUR business logic — your naming conventions, approval processes, workflow sequences — you are not automating generic tasks. You are building proprietary AI capabilities competitors cannot copy by buying a SaaS subscription.

Getting Started

You need three things:

1. An AI coding assistant: Cursor, Windsurf, or Claude Code

2. API access to the system you want to automate

3. A clear description of your workflow

Start with one system. Describe your workflow to the AI. Test it. Refine the instructions. Deploy it.

Then add a second system. Watch AI coordinate work across them like an experienced team member who never forgets a step.

The question is not whether your competitors will do this. The question is whether you will do it first.