Home › Fundamentals › What Is MCP Server Hosting? A Plain-English Explainer
What Is MCP Server Hosting? A Plain-English Explainer
MCP server hosting is the practice of running a Model Context Protocol server at a reachable address and operating it for you: HTTPS, authentication, scaling, credential storage and logging. You need it when anyone besides you, or any cloud-based assistant such as ChatGPT, must use the server. A local stdio server on your own machine needs none of it.
The short version of MCP
The Model Context Protocol is an open standard that Anthropic introduced in late 2024 for connecting AI clients to external tools and data. A server advertises tools, resources and prompts. A client such as Claude Desktop, Cursor or ChatGPT discovers them and lets the model call them. BusinessMCP’s own explainer on MCP uses the “USB for AI” comparison, which is a fair mental model: one plug shape, many devices.
The protocol defines two standard transports. stdio launches the server as a subprocess and talks over standard input and output. Streamable HTTP exposes the server as a network service. Hosting is a question that only the second one raises.
What hosting actually covers
When people say “hosted MCP server” they bundle several jobs together. Pulling them apart helps you judge any offer.
| Job | What it means in practice | Who owns it if self-hosted |
|---|---|---|
| Transport and TLS | A stable https URL speaking streamable HTTP | You |
| Authentication | OAuth or scoped bearer tokens on every request | You |
| Credential storage | Keeping upstream API keys and tokens out of agents’ reach | You |
| Scaling and uptime | Handling concurrent clients, restarts, deploys | You |
| Logging and audit | Recording which tool was called, with what, by whom | You |
| Protocol upkeep | Following new spec revisions and SDK changes | You |
A managed host takes some or all of those rows on. BusinessMCP’s remote versus local guide lists the same burdens for remote servers: hosting, uptime, TLS, monitoring and security. Those are exactly the items a host sells you relief from.
Key takeaway: the protocol is the easy part. Hosting is about operating a public, authenticated service that holds credentials and acts on behalf of an AI agent.
When you can skip hosting
Local servers have real advantages. The trust boundary is your operating system user, so there is no network auth to design. They can read local files, local databases and OS APIs. The cost is that each install serves one machine, and the server runs with your full privileges, so you should only install ones you trust.
Skip hosting when the tool is personal, when it must touch your filesystem, or when you are still prototyping. Move to hosting when any of these becomes true:
- A teammate needs the same tools without installing anything.
- You want a cloud assistant to use it. ChatGPT cannot launch local processes, so it needs a remote server.
- The server holds business credentials that should not sit on a laptop.
- You need an audit trail of who called what.
The trade-offs are covered in more depth in remote MCP server hosting and in self-hosted versus managed.
Three ways to get a hosted server
Run it yourself. Package your server, put it behind HTTPS on a container platform or serverless runtime, add auth and logging. Maximum control, maximum responsibility. The deploy walkthrough shows the steps.
Use a registry that also hosts. Some registries let developers publish servers that others install, and some run those servers for you. Smithery is one example, covered in our Smithery note.
Use a managed endpoint over your own tools. Here you do not write a server. You connect sources and the host exposes them as tools. BusinessMCP’s MCP hosting works this way: connect Stripe, ad accounts, Search Console, a CRM or a mailbox, and the workspace gets one endpoint at /api/mcp secured by a bearer key. The page also says it aggregates third-party MCP servers you connect, re-registering their tools under an ext_ prefix.
That last point comes with a limit the vendor states itself: it does not host your own MCP server code. If you wrote a server and want it run for you, a managed data endpoint is the wrong category.
What to ask before choosing
Whichever route you take, these questions separate serious offers from thin ones:
- Does it speak streamable HTTP and support OAuth as well as bearer keys?
- Can you issue one key per agent and revoke it instantly?
- Where do upstream credentials live, and can agents ever see them?
- Is every tool call logged?
- Can you export your data and leave?
- What happens when the spec revises? BusinessMCP says its endpoint follows the 2026-07-28 protocol revision and updates when that revision does; ask any vendor the equivalent.
BusinessMCP’s what is MCP hosting guide frames its own answers as an eight-point evaluation checklist. It is vendor-authored, so treat it as a useful list of questions rather than neutral proof. We expand it in MCP server hosting requirements.
A realistic starting point
Most teams do not need to decide everything at once. Keep the prototype local. When a second person needs it, put it remote with a bearer token and logging. When more than a few people or any sensitive data is involved, add OAuth, per-consumer keys and monitoring, or hand that layer to a host. Cost matters at each step, which is why we wrote what MCP hosting costs.
Common misunderstandings
“Hosting means the model runs there.” It does not. The MCP server exposes tools; the model runs wherever the AI client sends its prompts. You can even bring your own model key on some platforms. BusinessMCP says you can use your own Anthropic, OpenAI or Google key so calls go to your account.
“A hosted server is automatically safe.” A host can provide good defaults, but a poorly scoped tool is dangerous anywhere. The question is what each tool can do and which agent holds the key.
“One server per tool is the right design.” Usually the opposite. Fewer, well-described tools work better for models than many vague ones, and one endpoint per workspace is easier to govern than a dozen scattered ones.
| Misunderstanding | Reality |
|---|---|
| Hosting needs a new protocol | It reuses ordinary HTTPS plus MCP’s handshake |
| Local is deprecated | Local stdio remains the right choice for personal tools |
| Managed hosting runs any code | Some do, some, like BusinessMCP’s, only expose connected sources |
Frequently asked questions
What does it mean to host an MCP server?
It means running the server at a network address, usually an https URL, so AI clients can connect over the streamable HTTP transport, and operating it: authentication, TLS, scaling, credential storage and logging.
Does every MCP server need hosting?
No. A local server launched by your AI client over stdio runs as a subprocess on your machine and needs no hosting. Hosting matters when other people or cloud assistants must use it.
What is the difference between an MCP server and MCP hosting?
The server is the program that exposes tools and data. Hosting is the service around it that keeps it reachable, authenticated and monitored. You can host your own server or use a host that provides one.
Is MCP hosting the same as web hosting?
It overlaps, since remote MCP servers are HTTPS services. But MCP adds protocol handshakes, tool definitions and agent-specific auth concerns that ordinary web hosting does not address on its own.
Sources
- Model Context Protocol – Transports specification – https://modelcontextprotocol.io/specification/2025-06-18/basic/transports
- Model Context Protocol – Authorization specification – https://modelcontextprotocol.io/specification/2025-06-18/basic/authorization
- Anthropic – Introducing the Model Context Protocol – https://www.anthropic.com/news/model-context-protocol