MCP Server Development Service
Connect the systems you already run to AI agents through 1 MCP server.
We reply within 24 working hours. No commitment.
Swipe the diagram sideways
MCP Server Development: 4 existing systems connect to 4 AI clients through 1 MCP server
// problems we solve
Problems we often find.
A vendor offers a new system so an AI can read your data. Your ERP, spreadsheet, and CRM already run today.
We do not build from 0. 1 small MCP server reads and writes your existing systems through their API or database.
Staff copy data from the ERP into ChatGPT by hand. The data goes stale fast, and nobody knows what was shared.
A read tool fetches the data at the moment of the question. Every tool call lands in the audit log with the actor and the time.
You want an AI agent to draft invoices or change stock, but you fear a wrong write into production.
Every write tool uses human approval and an idempotency key. The agent proposes, your person approves, then the server writes 1 time.
Your team uses Claude, ChatGPT, OpenClaw, or Hermes Agent at the same time. A separate integration per client costs too much.
1 MCP server serves all of those MCP clients. We test with 1 client of your choice, and the other clients use the same server.
// how it works
The steps we run.
Swipe the diagram sideways
Source system audit
You share the API documentation or test database access. We map the data, pick the useful tools, and write the tool list with its access rights.
Output: A written tool list: name, purpose, read or write, and who approves.Build and test tools
We build the MCP server, set up an API key or OAuth 2.1, and test every tool with test data, not production data. Standard and Custom stretch this phase by tool count.
Output: Every tool passes the agreed test, with a record of the result.Connect the AI client
We connect the server to 1 AI client of your choice, turn on human approval for write tools, and install the audit log.
Output: Your AI client runs the read tools, and write tools wait for your person to approve.SOP and handover
We write the tool documentation, the usage SOP, and how to revoke access. A handover session runs with your team.
Output: Documentation, SOP, and credentials move to you.// what you get
A clear scope. Useful handover.
We build 1 small MCPMCPA standard that lets an AI program use your tools and data under clear permissions.Open the glossary server that connects your ERP, spreadsheet, database, or CRM to Claude, ChatGPT, OpenClaw, and Hermes AgentAI agentAn AI program that performs work steps by itself, for example reading a message, drafting a reply, and recording the result.Open the glossary.
- Source-system audit and a tool list
- MCPMCPA standard that lets an AI program use your tools and data under clear permissions.Open the glossary server with read tools and write tools
- APIAPIThe official door 2 systems use to exchange data, without anybody copying it by hand.Open the glossary key or OAuth 2.1, human approval, and audit log
- Test with 1 AI client of your choice
- Documentation, SOP, and handover
The result: 1 running MCP server (stdio or Streamable HTTP), a tool list with access rights, test results, documentation, and a usage SOP.
// price and inclusions
What the price covers.
| Package | Price | Scope | When to pick |
|---|---|---|---|
| Starter | IDR 2.5M, fixed price | 3–5 working days · 1 MCP server · 1 source system · up to 3 read tools · stdio or Streamable HTTP · API key · test with 1 AI client · documentation and SOP | Pick this when you want an AI agent to read 1 system first, with no write action. |
| Standard | IDR 6.5M, fixed price | 1–2 weeks · 1 server · up to 2 source systems · up to 8 tools including write tools with human approval · API key or OAuth 2.1 · hosting on Cloudflare Workers or your VPS · audit log · basic monitoring | Pick this when the agent must write to a system, for example an invoice draft or a stock change. |
| Custom | From IDR 12M, final price after review | 2–4 weeks · 3 or more source systems, or more than 8 tools · OAuth 2.1 or SSO · multi-tenant · rate limit · handover SLA | Pick this when more than 1 division or more than 1 company uses the same server. |
| Included | Not included | Billed apart |
|---|---|---|
|
|
|
Ground rules
- Revision: 1 round of fixes within 7 working days after handover, for the same tool list.
- Schedule: duration counts from complete source-system access, meaning API documentation or test credentials. Cancel before work starts at no cost.
- Ownership: the code, the server, and the credentials are yours. We remove our access when the 7 working day fix window ends.
- Warranty limit: we guarantee that the tools pass the agreed test, not a business result. Bugs in an AI client such as Claude, ChatGPT, OpenClaw, or Hermes Agent sit outside our control.
// is this for you
Good fit for
- You already run an ERP, a spreadsheet, a database, or a CRM.
- You want Claude, ChatGPT, OpenClaw, or Hermes Agent to read that data without copy and paste.
- You have API documentation or test database access.
Not a fit for
- You have no source system yet and need a new application. Choose Web Application Development.
- You have no agent yet and want it installed first. Choose AI Agent Ops Install.
What you prepare
- API documentation or test credentials for the source system.
- A list of 3 tasks you want the AI to help with.
- The name of the AI client your team uses.
What the price excludes
- Hosting, model API, and source-system licence fees.
- New tools added after handover.
How it differs: AI Agent Ops Install installs and hardens an agent, while MCP Server Development Service bridges your systems to the agent you already have.
// how this differs
Compare with other services.
| MCP Server Development Service | AI Agent Ops Install | Web Application Development | |
|---|---|---|---|
| Duration | 3 days–4 weeks | 1–2 weeks | 6–8 weeks |
| Price | From IDR 2.5M; Standard IDR 6.5M; Custom from IDR 12M | IDR 7.5–25M, final price after review | From IDR 25M, final price after review |
| Result | 1 MCP server that connects your systems to AI clients | An installed agent with access rights, an audit log, and an SOP | A new web application with a UI, an API, and a dashboard |
| When to pick | Your systems exist; you only need a bridge to an AI agent | You have no agent yet, or the agent needs installing and hardening | You need a new workflow that has no system yet |
// evidence and sources
Evidence we checked.
The MCP specification, revision 2026-07-28, defines 2 standard transports: stdio for a local server and Streamable HTTP for a remote server. The official schema holds 5 published revisions since 2024-11-05.
Model Context Protocol, Specification 2026-07-28: TransportsRead 13 September 2026The MCP tools specification asks for a human who can deny a tool call at all times. A server must validate input, control access, and limit the rate. A client should log tool use for audit.
Model Context Protocol, Specification 2026-07-28: ToolsRead 13 September 2026Cloudflare supports building and deploying remote MCP servers on Workers with Streamable HTTP and OAuth. Cloudflare states that the SSE transport for remote MCP is deprecated.
Cloudflare Docs, Model Context Protocol (MCP) on CloudflareRead 13 September 2026
- What Is MCP (Model Context Protocol)? An Explanation for Business Owners
- MCP Server Development Cost: The Scope Formula for Tools, Systems, Auth, Hosting
- MCP Server Security: Access Rights, Human Approval, Audit Log, OAuth 2.1
- Connect Legacy Systems to AI Agents Without a Rebuild: The MCP Pattern for ERP, CRM, and Spreadsheets
Owner: Rama Digital team · Updated 13 September 2026
// definitions
Terms we use.
- MCP
- MCP (Model Context Protocol) is an open protocol that connects an AI application to data and tools outside that application. The official site uses the USB-C analogy: 1 standard way for every connection. See in the glossary
- MCP server
- An MCP server is a small program that offers tools, resources, and prompts to an MCP client. It runs on your machine (stdio) or on the internet (Streamable HTTP).
- Read tool and write tool
- A tool is a function the AI model runs through the MCP server. A read tool only fetches data. A write tool changes data, so we put human approval on it.
- Human approval
- Human approval is the step where your person approves or denies 1 write tool call before the server runs it. The MCP specification recommends this step on every tool call.
FAQ
What is the MCP Server Development Service, and what does it cost?
MCP Server Development Service builds 1 MCP server that connects the systems you already run to an AI agent. The Starter package, IDR 2.5M, covers 1 source system and up to 3 read tools in 3–5 working days. Standard, IDR 6.5M, adds write tools with human approval, hosting, and an audit log. Custom starts at IDR 12M for 3 or more systems, OAuth 2.1 or SSO, and multi-tenant use.
Do we have to replace our systems so the AI can read them?
No. The MCP server reads and writes your existing systems through their API or database. Your ERP, spreadsheet, database, and WhatsApp CRM keep running as they do today. We only add 1 small program beside them. When a source system has no API, we check another path, such as a scheduled export or read-only database access, before we write the offer.
Which AI clients can use this MCP server?
We test the server with 1 AI client of your choice, and the other clients then use the same server. On 13 September 2026 we verified Claude (claude.ai, Claude Desktop, Claude Code), ChatGPT and Codex through a remote server, and Gemini API and Gemini CLI. OpenClaw, Hermes Agent, VS Code, and Cursor also support MCP servers. Each client has its own limits; for example, ChatGPT accepts a remote server only, not a local one.
How are access rights and data security handled?
Each tool has its own access rights: a read tool only reads, and a write tool needs human approval before the server runs it. A remote server uses an API key or OAuth 2.1 with a minimum scope. Every tool call lands in the audit log with the actor and time. Personal data columns can be filtered inside the tool, so data you do not need never reaches the AI model. We test with test data, not production data.
How long does it take to build an MCP server?
Starter takes 3–5 working days, Standard 1–2 weeks, and Custom 2–4 weeks. The clock starts when source-system access is complete: API documentation or test credentials are in our hands. Most time goes into understanding the source-system data, not into writing server code. Send the API documentation early, and the start date can be set sooner.
How is this different from AI Agent Ops Install and Web Application Development?
AI Agent Ops Install installs and hardens an AI agent on your server, with access rights and an SOP, from IDR 7.5M. Web Application Development builds a new application with a UI, an API, and a dashboard, from IDR 25M. MCP Server Development Service builds no application and installs no agent. It builds 1 bridge from your existing systems to the agent you already use.
Where is the MCP server hosted?
There are 3 options. A local server (stdio) runs on the user computer and suits 1 person. Cloudflare Workers runs a remote server (Streamable HTTP) from USD 5 per month on the Paid plan. Your own VPS also works, with the Origin check and the authentication we set up. Hosting fees go to the provider, apart from our price. The Standard package includes the hosting setup.
How do you calculate scope and price outside the packages?
Start from Starter at IDR 2.5M. Add-ons: read tool IDR 350 thousand each, write tool with approval IDR 600 thousand each, extra source system IDR 1.2M each. OAuth 2.1 or SSO adds IDR 1.5M, and hosting with monitoring adds IDR 900 thousand. Example: Starter + 2 read tools + 3 write tools + 1 system + hosting = IDR 7.1M. Standard at IDR 6.5M costs less, so the package wins. The final figure goes into the quotation.
Which system should we connect first?
Pick the system your staff ask about most often, and start with read tools. Examples that pay off early: stock checks in the ERP, document search in the drive, or order status in the WhatsApp CRM. A read tool changes no data, so the risk stays low and the result shows within days. Write tools, such as an invoice draft, follow once the team trusts the read tool answers.
Does Rama Digital run MCP servers of its own?
Yes. On our own systems, 3 MCP servers serve content, quotations, and scheduling for an OpenClaw agent and for Claude. Write actions use explicit confirmation, scoped access rights, and an audit log. We tested those servers with test data, not client data. We apply the same pattern to your server: small tools per goal, human approval on write tools, and records you can check.
// next step
Your systems stay. An AI agent reads them within 3–5 working days.
Send the source system name, 3 tasks you want the AI to help with, and the AI client your team uses. We reply with a proposed tool list and a price within 24 working hours.
