MCP Server Development Cost and 5 Pre-Commission Checks

MCP server development — what it means

MCP server development means building the connection point that lets AI operate your own systems and data directly. In Japan the outsourcing benchmark from the adjacent field of API integration development runs from ¥1,000,000 to over ¥10,000,000 (per GXO Inc., as of 2026). We build them at ¥2,500,000 for a read-only MCP Starter and ¥5,000,000 for an API + MCP Full Pack that includes authentication, rate limiting, audit logs, and published documentation (both excl. tax).

Hi, I’m Masato. I started as a Webflow specialist and have been widening my main field into AI website production and AI development. I run a web production company called Webharu in Shiogama, Miyagi, Japan.

“The board or a customer told us to make our SaaS MCP-ready.” “We want AI to reach our internal systems, but an accident would be terrifying.” This article is written for the CTOs, engineering leads, product owners, and IT managers of companies already operating an existing service — a SaaS, a business system. If you are an individual or a small team who wants to build something from scratch, this article is not aimed at you (we are preparing separate learning material for individual developers).

The conclusion first. MCP server development can be outsourced. On cost: putting comparable requirements to a major Japanese systems integrator runs ¥5,000,000–12,000,000 in total even for a small project (per SIA Inc., as of 2026). We take on the whole package at ¥2,500,000 for a read-only MCP Starter and ¥5,000,000 for the API + MCP Full Pack (both excl. tax). And I have implemented MCP servers on two of our own SaaS products (a CMS and an email API), both running in production. As someone on the building-and-operating side, I have pulled together the market data and the five things to check before you commission anyone.

The engineering-side design decisions — what to keep out of reach, how to design keys — are covered in detail in how we implemented an MCP server on our own CMS. This article stays on the buyer’s side of the table.

What an MCP server is: the connection point that lets AI operate your systems

An MCP server is a connection point that lets AI such as Claude or ChatGPT operate your own systems and data directly. Because it is built to a shared standard called MCP (Model Context Protocol), it is not locked to one AI service: any compatible AI client can use it.

As a buyer, there is only one thing to internalize. AI without an MCP server means a way of working where a human sits in front of a screen and copies and pastes as the intermediary. AI with an MCP server means a way of working where the AI operates your systems directly. The line between companies stuck at summarizing meeting notes and drafting emails, and companies where AI actually moves the work, is usually the presence of that connection point.

What MCP server development makes possible (three concrete examples)

Examples beat abstractions. All three are based on things we have actually built and are running.

1. Updating your CMS or business system from AI

You ask the AI, “turn last week’s trade-show report into an article and publish it,” and the AI creates the post directly in your CMS. That state is buildable. Our own CMS SaaS, HaruCMS, ships with an MCP server (eight tools) so posts can be created, updated, and illustrated with uploaded images straight from AI. It is not limited to CMSs — inventory management, customer management, any system where people type into a screen by hand fits the same shape.

2. Handing email over to an AI agent

The AI reads incoming mail, sorts it, and drafts the replies. Our email API, HaruMail, exposes sending and receiving to AI through an MCP server. The key part is misfire prevention: every sending tool can only go as far as creating a draft, and actual delivery can only be executed through a separate tool called confirm_send — a two-stage design. The line between what the AI handles and what a human keeps hold of is drawn in the architecture itself.

3. Letting AI read internal data safely

When companies want AI to reference internal documents and databases, most of them today paste into a chat window. Build a single read-only MCP server and the AI can fetch exactly the data it needs, by itself. Since you can configure it to permit no writes whatsoever, it suits a small first step: start by letting AI read.

The market cost of outsourcing MCP server development: a direct answer with data

First, the premise. As of 2026, MCP server development does not yet have an established market rate. In our own survey (August 2026), only four Japanese companies effectively publish MCP development as a productized service page, and only one publishes a price. So let me line up, with sources, the Japanese market rate for the adjacent field of API integration development, plus the US MCP rates that are running ahead of us.

Japan: API integration development runs ¥1,000,000 to over ¥10,000,000

• Per GXO Inc.’s market breakdown: ¥300,000–1,000,000 for a single API connection, ¥1,000,000–3,000,000 for multiple API integrations, ¥3,000,000–6,000,000 for workflow integration, and ¥6,000,000 to over ¥10,000,000 to API-enable a core system
• Per SIA Inc., contract development for large enterprises actually runs ¥1,000,000–1,800,000 per person-month, and even small projects total ¥5,000,000–12,000,000
• Person-month benchmarks are ¥800,000–1,200,000 at small and mid-sized development firms and ¥1,500,000–2,000,000 at major systems integrators (per CREX and Cross Network)

The US: enterprise-grade MCP runs past $100K

• Per Bacancy, MCP server development runs $18K–50K for an MVP and $110K to over $400K at enterprise grade
• Per Intuz, MCP integration for SaaS runs $10K–50K in a basic configuration and past $100K in an enterprise configuration
• Per Launch Day Advisors, marketplace-grade MCP servers run $100K–300K read-only and $300K–700K with writes

Maintenance runs ¥120,000–330,000 a month, and keeping up with API specification changes runs ¥100,000–300,000 per occurrence (per GXO Inc.).

Our pricing: Starter ¥2,500,000 / Full Pack ¥5,000,000 (excl. tax)

With that market data in mind, our pricing has three tiers (all excl. tax).

• MCP Starter (read-only) ¥2,500,000 — an MCP layer laid over your existing REST API. No writes, so accident risk is eliminated structurally. An entry point; roughly 4–6 weeks
• API + MCP Full Pack ¥5,000,000 — everything from REST API design through authentication and authorization (pk/sk separation, instant revocation on suspension), rate limiting, audit logs, two-stage handling of dangerous operations, the MCP server, and developer documentation plus llms.txt
• Full Integration from ¥10,000,000 — writes fully open, multiple AI clients, integration per business system. Quoted individually

Operations and maintenance start at ¥150,000/month (keeping up with API specification changes, monitoring, key management). Let me also state plainly where that sits against the market. Put the same requirements to a major systems integrator and over ¥10,000,000 is the going rate. We have already implemented and are running this same configuration on two of our own SaaS products, so we return the value of having a proven pattern to the price and deliver the whole package at ¥5,000,000. This is not a price to choose because it is cheap. Choose it after confirming that the pattern is already running in production.

The yardstick for comparing quotes: what is included, not the number

What to verify in an MCP server development quote is whether these four are included.

• Authentication — the mechanism that manages who may use it, by key. Does it cover the design of key issuance and revocation?
• Rate limiting — the mechanism that keeps the system from falling over when AI runs away with it or outside access spikes
• Audit logs — a full record of when, which key, did what. The foundation for incident investigation and internal controls
• Documentation — a specification both humans and AI can read. The asset that keeps it usable in-house after rollout

A bare working MCP server can certainly be built for less. But put one into business use without those four and the gaps erupt after go-live: a departed employee’s key still alive, an incident whose cause cannot be traced. Pricing and process details are collected on the API + MCP development pricing page (Japanese).

A five-item pre-commission checklist (usable as-is for comparing quotes)

This is the part I most want you to take away. When you request quotes from development firms, ask these five questions. All five are items we have confirmed we cannot do without while running our own SaaS in production.

1. Key revocation design: do a departed employee’s keys stop?

Ask: “when someone leaves or transfers, what happens to the API keys they issued?” If issued keys stay alive after their login account is disabled, you are left with a state where someone who left can still write to your system from outside. In HaruCMS, suspending a member automatically revokes the keys that person issued.

2. Is there a structural defense against mistaken operations?

Ask: “is the mechanism that stops AI from mass-sending an email or deleting data by mistake a matter of prompting, or of structure?” “We instruct the AI carefully, so it’s fine” is not an answer. HaruMail simply does not provide a tool that can send in one shot: delivery can only be executed through the two stages of draft creation and confirm_send.

3. Audit logs: is every operation recorded?

Ask: “can we trace afterwards when, which key, did what — all of it?” AI operations are numerous and fast, so with no record you cannot follow the cause when something goes wrong. We record every write operation across our own products in audit logs.

4. Is the tool design mindful of context consumption?

Slightly technical, but it hugely affects day-to-day usability. AI has a ceiling on how much information it can handle at once (its context), and an MCP server with too many tools or overly large responses eats up that capacity the moment you connect it. When I tried a general-purpose MCP from an external SaaS (Webflow MCP 1.0), merely connecting it loaded a huge amount of information and consumed most of the context — it never reached practical use (measurements in my write-up of testing Webflow MCP). That experience is why we keep tool counts tight and responses small, without exception. Try asking: “how will you limit the number of tools and the size of what they return?”

5. Is the documentation AI-readable?

Ask: “does the delivered documentation include a machine-readable form such as llms.txt?” Installing an MCP server and rolling it out internally is, these days, work the AI agents do themselves. With a spec an AI can read, connection is a matter of pointing the adopting side’s AI at it. For HaruCMS we publish both developer documentation and llms.txt.

Our track record: MCP servers implemented and running in production on two of our own SaaS products

Not “we can build it,” but “we built it, we sell it, and we run it.”

• HaruCMS (our own CMS SaaS, on sale since August 2026) — a REST API plus an MCP server (eight tools), with developer documentation and llms.txt published openly. Keys are split into pk (public key) and sk (secret key), and suspending a member automatically revokes that person’s keys. Every write operation is recorded in audit logs
• HaruMail (an email API, currently invitation-only early access) — an email API usable in Japanese, plus an MCP server. Actual delivery uses the two-stage confirm_send design, preventing AI misfires structurally

Beyond those, our own operations platform has an admin MCP (writes only within a safe allow-list, every operation recorded in audit logs) that we use internally every day. The substance of the design decisions is published in the article where I wrote up every design decision behind our MCP server implementation — useful material if your engineers want to scrutinize it.

The three steps from request to delivery

Step 1: A free consultation to judge whether AI can take this on

We hear what you want to do and judge whether that work can — and should — be handed to AI. A consultation for the judgment alone is perfectly fine. Whether your existing system has an API changes both configuration and cost, so we check the current state here too.

Step 2: Lock down the safety design, then implement

We settle the scope of key permissions, the protection against mistaken operations, and the audit log design first, and implement on top of that. Authentication and rate limiting are included. Building something that works and bolting safety on afterwards is not an approach we take.

Step 3: Deliver the documentation and connect to your AI client

Delivery includes both human-facing documentation and AI-readable documentation (llms.txt), and we confirm it working from the AI client you actually use (Claude, ChatGPT, and so on).

Summary: outsource MCP server development with the safety design included

The trick to not getting burned when outsourcing MCP server development is to commission “a package you can operate safely,” not “something that works.” The five-item pre-commission checklist — key revocation design, two-stage protection against mistaken operations, audit logs, tool design, and AI-readable documentation — works as-is as questions for any development firm.

To repeat, this article is aimed at engineering leaders at companies that already have a SaaS or business system. Our pricing starts at ¥2,500,000 (excl. tax) for the Starter, so this contract price band does not fit individuals or small teams looking to build something new (we are preparing separate learning material for individual developers).

Consultations with us are free. Even a request just to judge “can AI take over this task of ours?” is welcome.
→API & MCP Development service details
→Free consultation form

Related reading: if you run a SaaS, see how to make your SaaS MCP-ready; if you want the whole picture of connecting AI agents to internal systems, see AI agents and system integration.

FAQ

What is the market cost of outsourcing MCP server development?

On 2026 data, the adjacent field in Japan (API integration development) runs ¥300,000–1,000,000 for a single API connection, ¥1,000,000–3,000,000 for multiple API integrations, and ¥6,000,000 to over ¥10,000,000 to API-enable a core system (per GXO Inc.); contract work for large enterprises totals ¥5,000,000–12,000,000 even at small scale (per SIA Inc.). US MCP development runs $110K to over $400K at enterprise grade (per Bacancy). We take on MCP Starter (read-only) at ¥2,500,000, API + MCP Full Pack at ¥5,000,000, and Full Integration from ¥10,000,000 (all excl. tax). When comparing numbers, check that authentication, rate limiting, audit logs, and documentation are all in the quote.

How long does development take?

For a Starter configuration that lays a read-only MCP layer over your existing REST API, roughly 4–6 weeks. Full Pack and above build the connection point (the API) itself, which adds a design phase, so it varies with requirements. We present a schedule together with the quote after hearing your current state in a free consultation.

How much of the work do you cover?

Everything from designing and implementing the connection point (the API) through authentication, rate limiting, audit logs, the MCP server, published documentation (llms.txt included), and confirming the connection from the AI client you use. We also take on API-only or CLI-only requests.

Could the AI delete data or send email by mistake?

We prevent it structurally. Our standard design has four parts: narrow the permission scope of keys, revoke keys automatically when a member is suspended, split dangerous operations such as sending into two stages (in HaruMail, actual delivery is isolated in a separate tool called confirm_send), and record every write operation in audit logs. Our principle is a design that does not depend on operating rules or prompts.

Can we start small?

Yes. We recommend the read-only MCP Starter (¥2,500,000 excl. tax, roughly 4–6 weeks). If you only need AI to read internal data, it can be built with no write permission at all, so you can verify the benefit with accident risk eliminated structurally. We also welcome a free consultation just to judge whether a given task can be handed to AI.

Can we commission CLI or API work instead of MCP?

Yes. We select and combine the shapes that fit your existing operations — REST API, CLI tool, MCP server — to match your requirements. We have also developed our own mechanism for calling AI from local apps and internal tools without an API key (a local Claude CLI bridge) and used it on real projects.