Custom MCP servers · fixed scope

MCP server developmentfor your CRM, database and internal tools

We build custom MCP servers that let Claude, ChatGPT, Cursor or your own agent look things up in the systems your team already uses — and change them only after a person confirms. Everyone signs in with their own account, every tool has a scope, every call lands in an audit log.

prototype
from $800 · 1 week
production
from $2,000 · 2–3 weeks
clients
Claude · ChatGPT · Cursor · own agent
access
read-only → write with confirmation
fig. 1

§02Definition

What is an MCP server?

An MCP server is a small service that gives an AI assistant a fixed set of tools for one of your systems: “find a deal”, “show overdue invoices”, “create a task”. It speaks the Model Context Protocol, an open standard that Claude, ChatGPT, Cursor and many agent frameworks support, so one server works in all of them. The AI never sees a password: it can only call the tools you allow, with the rights of the person using it.

A custom MCP server fits your data model and your rules. Off-the-shelf connectors cover popular SaaS apps, but rarely your custom fields, approval steps or in-house API.

Works with

  • Claudedesktop, web, Claude Code
  • ChatGPTapps and developer mode
  • Cursorand other IDEs
  • VS Code + Copilotagent mode
  • Your own agentany MCP-capable framework
fig. 2
AI clientMCPMCP servertoolssearch · draft · updateresourcesschemas · docsauthOAuth · roles · logAPISQLyour system

Why now

One server, two new app directories

  1. ChatGPT opens app submissions. Its Apps SDK is built on MCP.

    OpenAI
  2. Upwork ships an MCP server for hiring from Claude, ChatGPT and Cursor.

    Upwork
  3. Claude Marketplace launches with 2,000+ connectors and plugins; submissions open on 25 Sep.

    gHacks, Unite.AI

§03Use cases

Connect Claude to your CRM — and ChatGPT to your company database

Your team stays in the AI chat it already likes. The server brings in live data and turns “please do this” into a checked action.

  1. Sales · CRM

    “Which deals in negotiation haven’t moved for 10 days?”

    crm.list_deals {stage:"negotiation", idle_days:10}

    Seven deals, largest first, with owners and last activity.

  2. Finance · accounting

    “Who has owed us money for more than 30 days?”

    acc.overdue {days:30}

    Totals by customer, plus a drafted reminder for each.

  3. Support · knowledge base

    “Can annual plans get a partial refund?”

    kb.search {q:"refund annual plan"}

    A quote from the policy, with a link to the exact section.

  4. Operations · tasks

    “Create a task for Anna to call the supplier tomorrow.”

    tasks.create {assignee:"Anna", due:"tomorrow"}

    The AI shows the task and waits for “Confirm” before creating it.

§04Systems

MCP server for HubSpot, Salesforce, QuickBooks or Postgres

Most projects start with one system and 3–5 tools. If the answers live in documents rather than tables, a knowledge-base assistant (RAG) can sit behind the same server.

SystemTypical toolsHow we connect
CRMHubSpot, Salesforce, Pipedrivecontacts, deals by stage, notes, tasksofficial REST API, OAuth app
AccountingQuickBooks, Xeroinvoices, payments, overdue balancesofficial API, read scopes first
DatabasesPostgres, MySQL, BigQuerysaved reports, safe queries with row limitsread-only replica or a restricted role
Docs & knowledgeGoogle Drive, Notion, Confluencesearch with source links, page draftsAPI plus a search index
ERP & back officeOdoo, NetSuite, in-house toolsstock, orders, statusesREST or GraphQL, or an adapter we write
SpreadsheetsGoogle Sheets, Excel onlineread ranges, append rowsSheets API, Microsoft Graph

§05Interactive

MCP server designer

Pick a system and an access level to see the tools we’d build, how the AI uses them and the price. Then send the config to the form.

01System

02Access level

mcp.config.yamltools: 4

# mcp.config · CRM · Read + drafts

server: crm-mcp

transport: streamable-http

auth: oauth2.1 · per-user

tools:

- name: crm.search_contacts

description: "Find a contact or company by name, email or phone"

scope: contacts:read

- name: crm.list_deals

description: "Deals by stage, owner and days without activity"

scope: deals:read

- name: crm.get_deal

description: "Deal card with history, amount and next step"

scope: deals:read

- name: crm.draft_notedraft only

description: "Draft a note or follow-up email; a person sends it"

scope: notes:draft

chat · demo data
  1. You

    Write Northwind a follow-up about the renewal.

  2. crm.draft_note {deal:"Northwind renewal"}draft only
  3. AI

    The draft is saved to the deal in your CRM. Mark can review it and send it from there.

estimate

Prototype
from $8001 week · the read tools
Production server
from $2,0002–3 weeks · + draft tools, OAuth, roles, audit log, hosting

Care plan from $250/mo

§06Security

Security model: the AI gets your permissions, never your passwords

An MCP server is a checkpoint, not a tunnel. It decides who may call which tool and keeps a record of every call.

permissions.matrixrole × toolsearchinvoicesdraftsupdatedeleteSales managerAccountantSupportOwner
  • allowed
  • after confirmation
  • not available

audit.log

  • 10:42:15annacrm.update_stageconfirmed
  • 10:42:18maxacc.find_invoices403 · role
  • 10:42:21olgaacc.find_invoices200 · 12 rows
  1. Sign-in per person

    OAuth 2.1: the AI works under each user’s own account and role. A sales manager can’t read payroll by asking Claude nicely.

  2. Least privilege

    Each tool has its own scope; read-only by default, write tools added one by one.

  3. Confirmation for writes

    The AI proposes the change, a person presses “Confirm”. Deletion isn’t exposed unless you ask for it.

  4. Audit log

    Who, which tool, which arguments, what came back — searchable and exportable.

  5. Secrets stay yours

    API keys live in your secret store, not in prompts or chat history.

  6. Prompt-injection tests

    Before launch we test the server with poisoned documents and tricky requests.

§07Delivery

Prototype in a week, production in three

Fixed scope, short loops: by the end of week one you test the server in your own AI client.

fig. 4
week 1week 2week 3thencall & accessprototypeproductionlistingcare plan30 min · API docs · sandboxread-only · 2–3 toolsOAuth · roles · confirm · log · hostingoptional · platform reviewmonitoring · updates≈ 3 weeks to productionweek 1week 2week 3thencall & accessprototypeproductionlistingcare plan30 min · API docs · sandboxread-only · 2–3 toolsOAuth · roles · confirm · log · hostingoptional · platform reviewmonitoring · updates≈ 3 weeks to production
  1. Day 1

    Call and access

    30 minutes on what your team asks most. You share API docs and a sandbox or a test copy.

  2. Week 1

    Prototype

    2–3 read-only tools for one system, running in Claude, ChatGPT or Cursor on your test data.

  3. Weeks 2–3

    Production

    Sign-in, roles, write tools with confirmation, audit log, hosting, tests and handover docs.

  4. Then

    Listing and care

    Optional ChatGPT app or Claude connector packaging; monitoring and updates on a monthly plan.

§08ChatGPT apps · Claude connectors

ChatGPT Apps SDK and Claude connector development

The same MCP server can ship as a ChatGPT app and as a Claude connector: the Apps SDK is built on MCP, and Claude plugins can bundle MCP connectors. ChatGPT has taken app submissions since 17 December 2025; Claude Marketplace opened on 23 September 2026 with 2,000+ connectors and plugins. We package the server, prepare the listing and go through review with you — approval is the platform’s call.

  • Apps SDK
  • remote MCP connector
  • plugin bundle
  • listing & review
fig. 5
mcp-servertools · auth · logChatGPTapp · Apps SDKsince 17 Dec 2025Claudeconnector · pluginsince 23 Sep 2026private by default

§09Pricing

MCP server development services: prices

Three fixed starting points. The final price goes into the contract after a 30-minute call.

  • MCP-P11 week

    Prototype

    from $800

    access
    read-only
    systems
    one
    tools
    2–3
    runs in
    your AI client, test data
    code
    in your repository
    Start with a prototype
  • MCP-P22–3 weeks

    Production server

    from $2,000

    access
    read, drafts, writes with confirmation
    sign-in
    OAuth 2.1, roles, scopes
    audit
    log of every call
    hosting
    your cloud or ours
    also
    prompt-injection tests, docs
    Plan a production server
  • MCP-C1monthly

    Care plan

    from $250/mo

    monitoring
    uptime and error alerts
    updates
    new MCP versions, API changes
    tools
    small additions each month
    hosting
    managed, if it runs with us
    support
    direct line to the developer
    Ask about care

You pay for Claude or ChatGPT subscriptions and model usage in your own accounts. Several systems or a public directory app are quoted separately.

§10Compare

MCP server, n8n/Zapier workflow or custom GPT actions: when to choose which

MCP server

AI asks, server acts

Who starts it
A person asks in chat; the AI picks the tool
Where it works
Claude, ChatGPT, Cursor, Copilot, own agents
Permissions
Per user and per tool, with an audit log
Best for
Questions and one-off actions on live data

n8n / Zapier workflow

trigger → fixed steps

Who starts it
An event or a schedule
Where it works
In the background, no chat needed
Permissions
Usually one service account per workflow
Best for
Repeatable processes: orders, sync, reports

Custom GPT actions

one GPT, one API

Who starts it
A person, inside one custom GPT
Where it works
ChatGPT only
Permissions
API key or OAuth for that GPT
Best for
A quick demo for one team in ChatGPT

Often you need both: an n8n workflow moves orders every night, and the MCP server lets a manager ask about them.

FAQ

Questions about MCP server development

Price, timing, security and ownership — the short answers.

How much does MCP server development cost?

A read-only prototype for one system starts at $800 and takes about a week. A production server with OAuth sign-in, roles, an audit log and hosting starts at $2,000 and takes 2–3 weeks. The care plan starts at $250 a month. Claude or ChatGPT subscriptions and model usage are paid separately, in your own accounts. The final price is fixed in the contract after a 30-minute call.

How long does it take?

About a week to a working prototype in your AI client, then 2–3 weeks to production. A listing in the ChatGPT app directory or Claude Marketplace adds the platform’s review time, which we don’t control.

What do you need from me?

A list of 10–20 questions or actions your team wants, API docs or admin access to a sandbox or test copy, the roles of the people who will use it, and where it should run. Plus someone who can answer questions for half an hour twice a week.

Is it safe to give an AI access to our CRM or database?

The AI never gets passwords or a direct database connection. It calls only the tools we expose, under each user’s own permissions. Read-only is the default, writes need a person’s confirmation, and every call is logged. Before launch we test the server against prompt injection. If your data must stay in your infrastructure, the server runs in your cloud.

Who owns the code?

You do. The code lives in your GitHub or GitLab from day one, and after payment all rights are yours, along with the deployment notes. No licence fees, no lock-in. We sign an NDA on request.

Which AI clients will it work with?

Any client that supports MCP: Claude (desktop, web and Claude Code), ChatGPT (developer mode and apps), Cursor, VS Code with GitHub Copilot, and agents you build yourself. Which subscription plans allow custom connectors differs between clients and changes often — we check yours on the first call.

Can you publish it in the ChatGPT app directory or Claude Marketplace?

Yes. The ChatGPT Apps SDK is built on MCP, and Claude plugins can bundle MCP connectors, so one server can feed both. We package it, prepare the listing and go through review together with you. Approval is up to OpenAI or Anthropic.

What if our system has no API?

We connect to a read-only database replica or to scheduled exports, or write a small adapter around your back office. Browser automation is the last resort, because it breaks whenever the site changes.

{ mcp.config }

Describe your systems — get a plan and a fixed price

Answer a few questions or paste the designer config. The developer who builds the server replies within one business hour.

Hire the MCP developer, not an agency

  • A senior developer from Ukraine with an Android background, building with Claude Code and MCP every day.
  • Not a big MCP server development company: one person from the first call to deployment.
  • EU and UK time-zone overlap. More on hiring a senior developer in Ukraine.

Myrado Studio is independent and not affiliated with Anthropic, OpenAI, Cursor or GitHub. Product names only describe compatibility.

01Your setup

Systems · pick any
AI clients · pick any
Access
Who will use it
Hosting
Specs, screenshots, photos, voice — drop them here or paste with Ctrl+V
    • Reply within one business hour
    • NDA on request
    • Code in your repository