A custom MCP server for one system costs $1,500 to $15,000 at firms that publish their prices in November 2026, and $15,000 to $80,000 or more once it has per-user sign-in, tools that change data, and several systems behind it. Hosting is small, usually $20 to $200 a month. The recurring cost that matters is upkeep: the MCP specification had four dated revisions between March 2025 and July 2026, and published maintenance plans run from $300 to $8,000 a month.
This guide puts nine published price lists side by side, explains which four decisions move the price, shows what a server costs to run, and covers the extra work when the server is meant to be an app inside ChatGPT. If you only need to know what MCP is, start with what the Model Context Protocol is and when you need it. If the AI should plan and act on its own across many steps, the bigger budget is the agent, covered in what an AI agent costs to build and run.
How much does MCP server development cost?
Prices below were read on November 11, 2026. Every number comes from a firm that sells this work, so read them as asking prices for a defined scope, not as market statistics.
| Firm (date) | Smallest package | Production package | Larger builds | Maintenance |
|---|---|---|---|---|
| NOLU Consulting | $1,500: up to 10 tools, API-key auth, about 5 business days | $3,000: up to 25 tools, OAuth 2.1, 7 to 10 business days | quoted separately; extra tools $100 each | $300 a month |
| MCP SI Labs | $490 setup and security audit of existing servers; from $1,900 for one system, up to 10 tools, 7 to 10 business days | from $5,900: several systems, OAuth, per-user permissions, audit log, 3 to 5 weeks | n/a | $390 a month |
| Fire In Belly | $1,500 audit (1 week) | from $7,500: pilot server live in 2 weeks | typically $15,000 to $40,000 | from $2,000 a month |
| Touchlane | $8,999 Essential, about 2 weeks | $12,999 Professional, about 3 weeks | $19,999 Enterprise, about 4 weeks | n/a |
| Innovation Insight | $8,000 to $15,000: one system, 8 to 15 tools, OAuth, 3 to 5 weeks | $15,000 to $28,000: several back ends, per-user permissions, 5 to 8 weeks | $25,000 to $40,000+: public multi-tenant server for a SaaS, 6 to 10 weeks | from $1,500 a month |
| Duskel (Sep 8, 2026) | $6,000 to $12,000: internal, read-only, 4 to 6 tools, 2 to 3 weeks | $15,000 to $30,000: single-tenant with a few writes | $30,000 to $75,000: customer-facing, multi-tenant, OAuth, writes, audit log, 6 to 10 weeks | 10% to 15% of build a year for spec and SDK changes |
| Reflekt Lab (Oct 26, 2026) | $2,300 to $4,700 for a half-day scoping | from $17,500 for a first workflow in production, 2 to 4 weeks | each further workflow cheaper | from about $470 a month plus $50 to $100 hosting |
| Phos AI Labs (Jul 31, 2026) | $5,000 to $15,000: read-only, one system | $15,000 to $40,000: 1 to 3 systems, OAuth | $40,000 to $80,000+: regulated data, custom access control | $2,000 to $8,000 a month |
| Fleurant AI (updated Oct 18, 2026) | 7 to 17 engineer-days, about $5,000 to $13,000 at internal US cost | 38 to 86 days, about $28,000 to $64,000 | 95 to 220 days, about $71,000 to $164,000 | 15% to 25% of build effort a year |
Two patterns stand out. First, the cheap end is real: a read-only server over one clean API, with up to ten tools and a shared key, is a one- or two-week job, and several firms sell it at a fixed $1,500 to $8,999. Second, the jump to five figures happens at the same point in every list: when each user signs in as themselves, when tools can change records, and when more than one system sits behind the server.
Fleurant's table is the only one priced in engineer-days, with a day rate built from US Bureau of Labor Statistics wage data (about $745 a day fully loaded). It is internal cost, before an agency's margin, which makes it a useful check on any quote: ask the vendor for the same line items in days.
Phos AI Labs also quotes MCP developer hourly rates in the US at $50 to $82 an hour. Treat it as one firm's figure; no public survey of MCP-specific rates exists yet.
What moves the price
Reflekt Lab, Duskel and Fleurant name the same four drivers, in roughly this order of impact.
| Decision | Cheaper | More expensive | Why |
|---|---|---|---|
| Who signs in | One shared key or service account | Each user with their own OAuth login and permissions | Per-user access needs OAuth discovery, token validation on every call and permission mapping. Duskel puts it at 1 to 3 weeks on a multi-tenant system. |
| Read or write | Read-only tools | Tools that create, update, send or delete | Duskel estimates a write tool at 2 to 3 times the cost of a read tool: confirmation before it runs, idempotency so a retry does not repeat the action, and a tested rollback. Fleurant gives 0.5 to 1 day per extra read tool and 1 to 3 days per write tool. |
| How many systems | One modern, documented API | CRM plus ERP plus spreadsheets, or a legacy system with no web API | Each system brings its own auth, limits and failure modes. A system with no API needs an adapter first, which Reflekt Lab says is often most of the work. |
| Where it runs | Local (stdio) on one machine | Remote (Streamable HTTP) for a team or for customers | Remote adds hosting, TLS, monitoring and a security review. |
Tool count matters less than people expect. Duskel's point: a REST API with 200 endpoints can become a six-tool server, because the model works better with a few tools that match whole tasks ("order status") than with one tool per endpoint. A quote that prices per endpoint usually means nobody designed the tools yet.
Where the hours go
The protocol part is small. Duskel writes that scaffolding a server in the official SDK and exposing four read tools takes two or three days with tests. Fleurant's line-item model shows where the rest goes on a production server (tier 2: remote, OAuth, five to eight tools including writes):
| Work item | Engineer-days (tier 2) |
|---|---|
| Discovery, tool design, schemas | 3 to 5 |
| Tool implementation | 6 to 15 |
| Upstream API and data integration | 4 to 10 |
| Authorization (OAuth, token validation, scopes) | 5 to 10 |
| Per-user permission mapping | 3 to 8 |
| Validation, rate limits, error handling | 3 to 6 |
| Logging, tracing, audit trail | 3 to 6 |
| Evals and agent-level testing | 4 to 8 |
| Security review and fixes | 3 to 8 |
| Deployment and CI/CD | 2 to 5 |
| Documentation and handover | 1 to 3 |
| Total | 38 to 86 |
Two lines in that table get cut first and cause the most trouble later. The first is response shaping: a tool that returns 4,000 rows or a 200-field object fills the model's context and makes answers worse. Duskel reports that cutting one client's default payload by 70% roughly halved token spend per conversation. The second is evals: a fixed set of real prompts that checks which tool ran with which arguments, so you notice when a model upgrade or a one-word change to a tool description breaks something.
What an MCP server costs to run
The server itself runs no model. The model lives in your users' Claude, ChatGPT or Cursor, so there is no AI bill on your side unless your own agent is the client.
- Hosting: Fleurant works out about $18 a month for an always-on AWS Fargate task (0.5 vCPU, 1 GB) plus about $3.65 for a public IPv4 address, and about $45 a month for an always-warm Google Cloud Run instance. A lightly used internal server on Cloud Run can stay inside the free tier. Duskel and Reflekt Lab quote $50 to $200 a month on their own hosting.
- Maintenance plans: $300 (NOLU), $390 (MCP SI Labs), from $470 (Reflekt Lab), from $1,500 (Innovation Insight), from $2,000 (Fire In Belly), $2,000 to $8,000 (Phos AI Labs) a month. The cheap plans include a few hours of changes; the expensive ones include incident response and new tools.
- Tokens on the client side: every tool definition loads into the model's context before the first question. Duskel measured 4,000 to 6,000 tokens of preamble for 20 tools with long descriptions. Your users or your own agent pay for that in every conversation, which is why a short tool list is also a cost decision.
The maintenance line most quotes skip
The current MCP version is 2026-07-28. Earlier dated revisions were 2025-03-26, 2025-06-18 and 2025-11-25. The July 2026 revision removed protocol-level sessions and the initialize handshake: every request now declares its own protocol version, and servers must answer a server/discover request. It also deprecated OAuth Dynamic Client Registration in favour of Client ID Metadata Documents. Deprecated features stay in the spec for at least twelve months, or ninety days under an expedited exception.
In practice, a server built in 2025 and left alone will drift out of step with the clients, and a server built today may need to support both the old handshake and the new per-request versioning while clients catch up. Fleurant budgets 2 to 8 engineer-days per breaking revision for a production server; Duskel budgets 10% to 15% of build cost a year just for spec and SDK changes. If a quote has no maintenance line, ask who updates the server after the next revision and at what price.
Building an MCP server for ChatGPT
An app in ChatGPT is an MCP server with extra parts. OpenAI's quickstart describes three: the MCP server with tools, an optional web component that ChatGPT renders in an iframe (ChatGPT implements the open MCP Apps UI standard, so the same UI can run in other compatible hosts), and the metadata the model uses to decide when to call your tools. During development you connect it in ChatGPT as a custom MCP server over HTTPS.
Publishing in the ChatGPT directory adds review work that a private connector does not need. From OpenAI's current plugin guidelines:
- Every tool must set
readOnlyHint,destructiveHintandopenWorldHintto explicit true or false values. Anything that sends, uploads, cancels or deletes counts as a write, and being able to undo it does not make it non-destructive. - Each operation the model can call must be its own reviewable tool. A generic "execute any operation" tool is not allowed.
- A published privacy policy is required, covering what personal data you collect, why, who receives it, how long you keep it and what controls users have.
- Tools may only request the data needed for the task. No conversation history, no precise location.
- The app must work in ChatGPT on desktop and mobile, including the UI, and an authenticated server needs a working test login for reviewers.
- OAuth 2.1 is needed only for user-specific data or write actions; OpenAI's own guide suggests keeping read-only and anonymous flows simple.
What does that do to the price? None of the nine firms above publishes a separate ChatGPT-app price. Touchlane's packages explicitly target ChatGPT, Claude, Cursor and Perplexity, and include registry publication. A reasonable way to estimate it yourself: take the server tier your scope falls into, then add the widget (a small React or HTML front end), the privacy policy, annotation review and test cases for the submission. A read-only app with no UI is close to the cheap tier; a signed-in app with a custom UI and write actions sits in the production tier.
When you should not build one
- An official server already exists. GitHub, Salesforce, QuickBooks Online and many SaaS tools ship their own. Phos AI Labs and Reflekt Lab both say to check registries and vendor servers first and build only for the actions they lack.
- One person, one read-only tool. A local server over one API is a short job with the official SDKs. Pay someone when it has to serve a team, sign people in or write.
- Only one agent will ever use the integration. Fleurant's rule: if you control both ends and there is one consumer, a direct function call is less code. Add MCP when the second client appears.
- You do not know yet what the AI should do. Then spend on scoping, not code.
Freelancer, agency or in-house?
| Option | Fits when | Watch out for |
|---|---|---|
| Fixed-price specialist (NOLU, MCP SI Labs) | One SaaS API, clear tool list, you want it in 1 to 2 weeks | Scope caps: tool limits, what counts as a change, who hosts |
| Freelancer | You want one person who designs the tools, builds the server and talks to your team; a pilot first, then more systems | Ask what happens if they are unavailable, and get the code in your repository from day one |
| Agency | Regulated data, several systems, security review and on-call support in one contract | Higher price for the same tier; check who on their team has shipped a remote server with OAuth |
| In-house | MCP will be a permanent part of your product | Phos AI Labs puts a first production server at 1 to 3 months for a team without MCP experience |
A common pattern Phos AI Labs describes: an outside team builds the first server with documentation, then your engineers own the next iterations.
How to keep the cost down
- Start read-only, with one system and five or six tools that match real tasks.
- Log every tool call from day one (tool, arguments, latency, result size). After a month, delete tools nobody calls and rewrite the ones with high error rates.
- Add write tools one at a time, each with a confirmation step and a way to undo it.
- Write error messages for the model: "invalid customer_id, call find_customer first" gets fixed on the next turn, "400 Bad Request" gets retried unchanged.
- Get the code in your repository and the server in your cloud account, so a maintenance plan is optional, not a lock-in.
Questions to ask before you sign
- Which tools read and which write? What is the confirmation and undo path for each write?
- Does each user sign in with their own permissions, or does the server use one shared account?
- Which MCP revision do you target, and will older clients still connect?
- Which clients will you test against (Claude, ChatGPT, Cursor, your own agent)?
- How many tools are in scope, and how many tokens do their definitions use per conversation?
- Is a security review included? Who checks token audience validation and prompt injection through tool results?
- What does maintenance cost after launch, and what does it include?
- Who owns the code and the hosting account at the end?
FAQ
How much does it cost to build an MCP server?
At firms with public prices, $1,500 to $15,000 for one system, depending on tool count and sign-in (NOLU, MCP SI Labs, Touchlane, Innovation Insight, Duskel). Production servers with OAuth, write tools and several systems run $15,000 to $80,000 or more (Duskel, Phos AI Labs, Fire In Belly). Hosting adds $20 to $200 a month.
How long does it take to build an MCP server?
A read-only server for one API takes 1 to 3 weeks at specialist firms. A production server with per-user OAuth and write tools takes 3 to 10 weeks. Teams without MCP experience building in-house need 1 to 3 months for the first production server, according to Phos AI Labs.
How much does it cost to build an app for ChatGPT?
No firm in this comparison publishes a separate price. An app in ChatGPT is an MCP server plus an optional UI widget, a privacy policy and directory review work, so start from the server tier your scope fits and add the UI and submission work.
What does an MCP server cost per month?
Compute is about $18 to $45 a month on AWS Fargate or Google Cloud Run at published prices, often less with scale-to-zero. Maintenance plans run from $300 to $8,000 a month. The server runs no model, so there is no AI bill unless your own agent is the client.
Does an MCP server need OAuth?
Only if it is remote and acts for individual users or changes data. A local server, or a read-only server over public data, can use simpler credentials. For a team or for customers, per-user OAuth is the main cost driver in every price list above.
Why do MCP server quotes vary so much?
Because "an MCP server" covers anything from one read-only tool on a laptop to a multi-tenant service with OAuth, writes and an audit trail. Compare quotes on the same four points: who signs in, which tools write, how many systems, and where it runs.
Do I need a custom MCP server if my tools already have one?
Usually not. Check the vendor's own server and public registries first, and build only for the actions or the internal system they do not cover.
Need an MCP server or a ChatGPT app built?
I build AI integrations as a freelancer, including MCP servers. My own Business Ideas knowledge base runs as a public remote MCP server that Codex and Claude Code connect to over HTTP. I start with a workshop to pick the one workflow the server should handle, then a written scope and two-week sprints, with a working server your team can test after the first one. Projects start at $5,000, with no upper limit, because the price depends on scope. I bill hourly on a B2B invoice. See the AI integration service or AI agent development, or tell me which system the AI should connect to and I'll reply within 24 hours.
Sources
Read November 11, 2026, unless noted.
- NOLU Consulting, MCP servers for SaaS APIs
- MCP SI Labs, MCP server development pricing
- Fire In Belly, Custom MCP server development
- Touchlane, Custom MCP server development
- Innovation Insight, MCP server development services
- Duskel, What an MCP server actually costs to build (September 8, 2026)
- Reflekt Lab, How much does it cost to build an MCP server? Our real prices (2026) (October 26, 2026)
- Phos AI Labs, When to hire MCP server development services (updated July 31, 2026)
- Fleurant AI, MCP server development cost: build, run and maintain in 2026 (updated October 18, 2026)
- Model Context Protocol, Versioning and Key changes in 2026-07-28
- OpenAI, MCP server and UI quickstart, Authentication and Plugin guidelines
- OpenAI, Bring your app to ChatGPT
Want to build an AI agent, an MVP, or an app?
I design and build AI products from the first version to production: MVPs, React Native mobile apps, web apps, and AI agents. Tell me what you want to build and I will tell you where to start.