How much does an MCP server cost?
An MCP (Model Context Protocol) server lets an AI app such as Claude or ChatGPT use one of your systems through a short list of tools. With us, a read-only starter server with up to 3 tools is $1,999. A full server with its tools costs $3,500 to $7,500, fixed before the build. Bigger agent platforms are quoted per app.
By Babar Ali, founder · updated 15 November 2026 · outside prices and facts read on 14 and 15 November 2026, sources below
What moves the price of an MCP server?
The protocol is the cheap part. Seven decisions set where a server lands: a $1,999 starter, a full server inside $3,500 to $7,500, or a scoped build beyond that range. Most of them are about who may do what, not about code, which is why the first week of every engagement is a written tool map rather than programming. The MCP development service describes that work in full; this page is about what it costs.
| What moves it | Lower in the range when | Higher when |
|---|---|---|
| Number of tools | Three to five tools, each one clear task | Ten or more, or one tool per API endpoint |
| Sign-in | One API key or service account | OAuth, so each person signs in as themselves |
| Reads and writes | Read-only tools | Tools that create, change or delete records |
| Approval steps | Nothing writes, so nothing needs approving | A person approves each write before it runs |
| Hosting | Local, on each person's machine | Remote, in your cloud, shared by a team or by customers |
| Testing with real clients | One AI app your team already uses | Several AI apps, each run through real tasks |
| Security review | Internal data and a small, trusted team | Customer data, money, or content from outside the company |
The number of tools
Each tool is one task the model can ask for, with a name, a description and typed inputs. Designing and testing a tool costs more than writing it, so the count moves the price more than the lines of code do. The cheapest good server has a few tools shaped around what people actually ask, not one tool per endpoint of your API. The count also moves your running costs, because tool definitions travel with every request, as the running costs below show.
Sign-in: API keys or OAuth
The MCP specification makes authorization optional. When a remote server does sign people in, the specification’s flow is built on OAuth 2.1, and a local server should take its credentials from its environment instead[1]. Claude, for one, can connect to a custom server with each person signing in through the server’s OAuth flow, or with an API key sent as a request header, which its documentation lists as a beta for a limited set of organisations[2].
An API key is quicker to build. OAuth is right for anything shared, because each person keeps their own permissions, and it costs more: consent, token refresh, and never forwarding a person’s token to your other systems, the “token passthrough” that MCP’s security guidance says its authorization specification forbids[3].
Read tools, write tools and approval steps
Read tools are the cheap half. A write tool needs validation, a record of what changed and a decision about who must say yes first, so we put an approval step in front of every write. AI apps ask before calling tools too, but Claude’s documentation shows those prompts can be set to Always allow[2], so for a write that matters, the check belongs on the server. Vendors split the two as well: Linear’s hosted server gives read-write access by default and a separate read-only address that only ever exposes read tools[4].
Hosting: remote or local
MCP has two standard transports. Over stdio, the AI app starts the server as a subprocess on the same machine; over Streamable HTTP, each message is an HTTP POST to one server address[5]. A local server is quicker to build and has no hosting bill, and it suits a few developers. A remote server needs hosting and an HTTPS address the AI app can reach[2], and anything shared should sign people in. It is the only sensible shape for a team, several teams or your customers.
Testing with real clients
The MCP Inspector is the project’s reference tool for testing and debugging servers[6]. It is useful, but it is not the AI app your people will use. Clients differ in how they ask for approval, handle long tool lists and show results, so our test plan names each app your team or your customers use and runs real tasks through every one, including requests that should be refused. Each extra client adds test time.
The security review
MCP’s own security guidance covers attacks such as the confused deputy problem, token passthrough and server-side request forgery[3]. Prompt injection is the other risk: anything a tool returns, such as an email body or a web page, can carry instructions, and Claude’s documentation tells people adding a custom server to be aware of it[2]. We test every server with hostile content before handover. A formal penetration test is a separate engagement, under cybersecurity.
Want a fixed price for your server?
Send the system and the tasks you have in mind. The first week maps the tools and the permissions, and the price is fixed in writing before the build starts.
What does an MCP server cost at three sizes?
A worked example. The sizes are illustrations rather than packages: the tool map in the first week says which one you are in, and the price is fixed in writing before the build starts.
| Server size | What it includes | Price and time with us |
|---|---|---|
| Starter: read-only, up to 3 tools | Up to three read tools over one system with a clean API. One API key or service account. Runs locally over stdio, or remotely for a small team. Tested in the AI app your team uses, then handed over with documentation. | $1,999, live in about 2 weeks. A fourth tool, or any write, makes it a full server. |
| Medium: reads, writes and approvals | Six to ten tools, some of them writes, each write behind an approval step. People sign in as themselves with OAuth. Remote over HTTP in your cloud, with a log of every call. Tested in two or three AI apps, including the requests that should be refused. | The middle to the top of $3,500 to $7,500. Usually about three weeks. |
| Large: your customers, many tenants | A remote server your own customers sign in to, each seeing only their own data, with tools set by plan, rate limits and support for the AI apps your customers use. | Quoted per app, like other AI builds. Two weeks of scoping come first and end in a fixed quote in writing. |
Between the starter and the medium size sits a read-only server with four or more tools, at the low end of $3,500 to $7,500.
The jump from medium to large is not the tool count. Tenancy is: once your customers sign in, every tool has to know whose data it is touching, the server has to survive their traffic, and somebody has to support the AI apps they bring. That is why it is quoted like the AI agent builds it usually becomes part of: per app, fixed in writing once it is scoped, and below what US and EU agencies charge for the same build. How AI builds are quoted is on the pricing page.
What MCP studios charge, November 2026
Three studios publish what an MCP server costs, and Upwork publishes ranges for freelancers. We read each page on 15 November 2026. Their prices are in the table, each with its source, and ours follow it.
| Who | Published price | What the price is for | Source, and the day we read it |
|---|---|---|---|
| AppUnitePoznań, Poland | $2,500 fixed fee | Its own fixed fee for an MCP server for a SaaS product, sold as “Agent Ready in 14 Days” | AppUnite, MCP server implementation. No date shown. Read 15 November 2026. |
| RaftLabsoffices listed in Dublin and Ahmedabad | From $10,000. Its guide: $5,000 to $15,000 for a basic server, $15,000 to $40,000 for production | Its own starting price for a focused server. In the guide, a basic server has 3 to 5 tools and authentication; production adds OAuth 2.0 and several tool categories | RaftLabs, MCP server development, no date shown, and its guide to MCP companies in 2026, updated 30 October 2026. Both read 15 November 2026. |
| Groovy Weboffices listed in India, Australia, Germany and the US | $8,000 to $18,000 for a single server; $15,000 to $30,000 for a multi-tool server | Its guide: a single server connecting one internal tool, 2 to 4 weeks; a multi-tool server, 4 to 6 weeks | Groovy Web, MCP development cost, updated 23 October 2026. Read 15 November 2026. |
| Upwork’s guidefreelancers worldwide, not a studio | $1,500 to $5,000 for a single connector; $5,000 to $15,000 for a multi-tool server with auth | A server for one data source or API; a multi-tool server with OAuth, per-user scoping and audit logging | Upwork, hire MCP server developers. No date shown. Read 15 November 2026. |
With us: a starter server, read-only with up to 3 tools, is $1,999. A full server with its tools is $3,500 to $7,500.
AppUnite’s $2,500 fixed fee is close to our $1,999 starter, and both take about two weeks. Compare what each price includes before you choose on it: our starter is read-only with up to 3 tools. Upwork’s ranges are for freelancers, not studios, and a single connector there can cost less than either. Pages that show no date can change, so check the source before you quote it.
What does an MCP server cost to run each month?
Hosting is usually small. Model tokens are the line to watch, and only when your own agent is the one calling the model. These are the providers’ own published prices as of November 2026.
| Cost line | Published price, as of November 2026 | What it means for a server |
|---|---|---|
| A local server | No hosting: the AI app starts it on each person’s machine[5] | Free to run. Each person installs and updates it. |
| Cloudflare Workers | Free plan: 100,000 requests a day. Paid: $5 a month minimum, with 10 million requests included, then $0.30 per extra million[12] | 100,000 a day is about 3 million requests a month before anything is billed. |
| AWS Lambda | Free tier: one million requests and 400,000 GB-seconds a month, then $0.20 per million requests plus compute time[13] | Fits a server that already lives in your AWS account. |
| Google Cloud Run | Free tier with request-based billing: 2 million requests, 180,000 vCPU-seconds and 360,000 GiB-seconds a month[14] | Fits a server that already lives in Google Cloud. |
| Claude plans, when people use the Claude app | Free, Pro $20 a month ($17 billed yearly), Max $100, Ultra $200[15]. Every plan can add a custom server by URL; Free can add one[2] | Flat plans with usage limits, paid per person. No per-token bill to you. |
| A model API, when your own agent is the client | Claude Sonnet 5.5: $2 per million input tokens, $10 per million output. Claude Haiku 5.5: $0.10 and $0.50 for prompts up to 100,000 tokens[16]. OpenAI gpt-6.1-sol: $2 and $10 at standard rates[17] | Every request pays for the tool definitions as well as the conversation. |
On a model API, tool names, descriptions and input schemas count as input tokens on every request, and Claude adds its own tool-use system prompt, 286 tokens on Sonnet 5.5[18]. Say your ten tool definitions come to 2,000 tokens: each request then carries about 2,300 extra input tokens, roughly $4.57 per thousand requests at Sonnet 5.5’s input price[16], before the conversation itself and whatever the tools send back.
That arithmetic is one more reason to keep tool lists short and results small. The larger running cost is usually people rather than requests: someone has to rotate credentials, update the server when the system behind it changes, and read the log when an answer looks wrong. Plan for that time when you compare building against buying.
Start with a $1,999 starter server
Read-only with up to 3 tools, live in about 2 weeks. If an existing server would do the job, we will say so on the first call.
Should you build an MCP server or use one that exists?
Check for an official or open-source server before you pay anyone, including us. As of November 2026, these companies publish their own MCP servers:
- GitHub: an official, open-source server under the MIT licence, hosted by GitHub as a remote server or run locally[19]
- Stripe: an MCP server for the Stripe API and its documentation, in public preview[20]
- Notion: Notion MCP, for clients such as Claude, ChatGPT and Gemini[21]
- Linear: a centrally hosted server over Streamable HTTP with OAuth 2.1 sign-in[4]
- HubSpot: a remote server with read and write access to CRM data[22]
The Official MCP Registry lists many more[23]. Be careful with the reference servers in the MCP project’s own repository: the project calls them educational examples, not production-ready solutions[24]. And a server from someone you do not know is code you are trusting with your data. Claude’s documentation says custom servers are unverified and advises connecting only to trusted organisations[2].
An existing server is enough when
- The system is a mainstream SaaS tool with an official server
- Its tools match what your people actually ask for
- Everyone signs in as themselves, and its permission scopes suit you
- You are comfortable with what it lets the AI change
- It is open source, and someone on your team will read its code and keep it updated
Build your own when
- The system is yours: an internal API, a database or an old tool nobody else will cover
- You want a few task-shaped tools instead of one per endpoint
- Writes need your approval rules and your own log of every call
- Your customers should reach your product from their own AI apps
There is a middle path: run the official server read-only, and build only the one or two tools it lacks, with the approval step around them. It is often the cheapest honest answer.
When not to hire us
- An official server already does the job. Use it, and keep the money.
- You need a one-off data pull. An export or a script is cheaper and safer.
- It is two or three read-only tools and you have a developer with a free week. Build it, test it with the MCP Inspector, and keep the $1,999.
- Your budget is under $1,999. That is our floor, for a starter server, and a freelancer or your own team is the better trade.
- You want an AI to write to money, payroll or production data with no person approving it. We will not build that.
If the job is moving data between systems on a schedule, rather than giving an AI app new tools, you do not need an MCP server at all. AI automation is the cheaper and more honest answer, and the n8n, Make and Zapier comparison helps you pick the platform before anyone writes code.
We say the same on the first call. A server nobody needs is still a server somebody has to secure.
Free download · PDF, 3 pages · no email needed
MCP server prompt injection test checklist
The 12 prompt injection cases every server we build goes through, each with its safe result, and the 11 checks to run before you ship.
Download itQuestions about MCP server cost
With us, a starter MCP server, read-only with up to 3 tools, is $1,999 and live in about 2 weeks; a full server with its tools is $3,500 to $7,500, usually live in about three weeks. Every price is fixed in writing before the build starts. Read-only means every tool only reads data; any write tool needs the full server. A larger agent platform, with several servers, clients or tenants, is quoted per app like our other AI builds: scoping comes first and ends in a fixed quote in writing. Because the team is in Lahore, Pakistan, that quote comes in below what US and EU agencies charge for the same build.
About 2 weeks for a starter server, read-only with up to 3 tools. A full server usually takes about three weeks: the first week maps the tools and the permissions, then comes the build and the tests in the AI apps you use. Per-person sign-in, write tools with approval steps and several tenants take longer, and the tool map tells you which case you are in before you commit.
Often little or nothing. A local server runs on each person's own machine. Hosting a remote one costs little: as of November 2026, Cloudflare Workers' free plan allows 100,000 requests a day and its paid plan starts at $5 a month with 10 million requests included, and the AWS Lambda free tier includes one million requests a month. The bigger running cost is the time to keep the server current as the systems behind it change.
Only when your own agent calls a model API. Then every tool's name, description and input schema is sent as input tokens on each request, so a long tool list costs money on every call. When people use the server from the Claude app, the model runs on their own Claude plan, which is free or a flat monthly price, with usage limits, rather than a per-token bill.
Often, yes. As of November 2026 GitHub, Stripe, Notion, Linear and HubSpot publish their own MCP servers, GitHub's as open source under the MIT licence, and the Official MCP Registry lists many more, so check there first. The reference servers in the MCP project's own repository are a different thing: the project describes them as educational examples, not production-ready solutions.
An API key is one shared credential, so the server acts as a single service account. OAuth signs each person in as themselves, so the server has to respect each person's permissions, handle consent and refresh, and never pass a person's token straight through to your other systems, which the MCP authorization specification forbids. It is the right choice for anything shared, and it is more work.
The written tool map, the build, typed tools with reads and writes kept apart, an approval step in front of every write, a log of every call, tests in the AI apps you name, including requests that should be refused, and the handover: repository, deployment, documentation and 30 days of support. The server runs in your cloud account and you own the code. A starter server is read-only, so it has no write tools and no approval steps to build.
More tools, tools that write, approval steps, per-person sign-in, more AI apps to test in, and data that comes from outside the company, such as emails or web pages that could carry injected instructions. A multi-tenant server for your own customers is quoted per app, like an AI build, because tenancy and support decide its price more than the tool count does.
Sources
Every outside fact on this page comes from the pages below, and each entry gives the day we read it. Our own prices are the ones published on our pricing page.
- Model Context Protocol specification, version 2026-07-28: Authorization. modelcontextprotocol.io/specification/2026-07-28/basic/authorization. Read 14 November 2026.
- Claude documentation: Add a connector that isn't in the directory. claude.com/docs/connectors/custom/remote-mcp. Read 14 November 2026.
- Model Context Protocol documentation: Security best practices. modelcontextprotocol.io/docs/2026-07-28/tutorials/security/security_best_practices. Read 14 November 2026.
- Linear documentation: MCP server. linear.app/docs/mcp. Read 14 November 2026.
- Model Context Protocol specification, version 2026-07-28: Transports. modelcontextprotocol.io/specification/2026-07-28/basic/transports. Read 14 November 2026.
- Model Context Protocol documentation: MCP Inspector. modelcontextprotocol.io/docs/2026-07-28/tools/inspector. Read 14 November 2026.
- AppUnite: MCP server implementation (no date shown). www.appunite.com/main-services/mcp-server-implementation. Read 15 November 2026.
- RaftLabs: MCP server development (no date shown). www.raftlabs.com/services/mcp-server-development. Read 15 November 2026.
- RaftLabs blog: Top MCP development companies in 2026, by Riya Thambiraj, updated 30 October 2026. www.raftlabs.com/blog/top-mcp-development-companies. Read 15 November 2026.
- Groovy Web: MCP development cost guide, updated 23 October 2026. www.groovyweb.co/blog/mcp-development-cost. Read 15 November 2026.
- Upwork: MCP server developers hiring page (no date shown). www.upwork.com/hire/mcp-server-developers/. Read 15 November 2026.
- Cloudflare Workers documentation: Pricing. developers.cloudflare.com/workers/platform/pricing/. Read 14 November 2026.
- AWS Lambda pricing. aws.amazon.com/lambda/pricing/. Read 14 November 2026.
- Google Cloud Run pricing. cloud.google.com/run/pricing. Read 14 November 2026.
- Claude plans and pricing. claude.com/pricing. Read 14 November 2026.
- Claude API documentation: Pricing. platform.claude.com/docs/en/about-claude/pricing. Read 14 November 2026.
- OpenAI API documentation: Pricing. developers.openai.com/api/docs/pricing. Read 14 November 2026.
- Claude API documentation: Tool use overview, the Pricing section. platform.claude.com/docs/en/agents-and-tools/tool-use/overview. Read 14 November 2026.
- GitHub: github/github-mcp-server, GitHub's official MCP server. github.com/github/github-mcp-server. Read 14 November 2026.
- Stripe documentation: Model Context Protocol (MCP). docs.stripe.com/mcp. Read 14 November 2026.
- Notion developer documentation: Notion MCP overview. developers.notion.com/guides/mcp/overview. Read 14 November 2026.
- HubSpot developer documentation: HubSpot MCP server. developers.hubspot.com/ai-tools/mcp. Read 14 November 2026.
- Official MCP Registry. registry.modelcontextprotocol.io/. Read 14 November 2026.
- GitHub: modelcontextprotocol/servers, the MCP reference servers. github.com/modelcontextprotocol/servers. Read 14 November 2026.
Keep reading
Have a system your AI apps should be able to reach?
A starter MCP server, read-only with up to 3 tools, is $1,999 and live in about 2 weeks; a full server with its tools is $3,500 to $7,500, usually live in about three weeks. Send the system and the tasks you have in mind, and we reply within one business day with where we would start.