Agents · Updated 10 November 2026 · 8 min read

What is MCP, and how do hosting platforms use it?

The Model Context Protocol in plain words, the two ways hosts use it, and what to check before an agent touches your deployments.

View as Markdown

MCP, the Model Context Protocol, is an open standard that lets an AI application such as Claude Code, Codex, Cursor or ChatGPT call tools in another system. Hosting platforms use it in two ways: they publish their own MCP server so your coding agent can create projects, deploy and read logs for you, and some of them also run MCP servers that you write. Eurobase, Vercel, Render, Railway, Netlify and Cloudflare all publish a server of the first kind, and Vercel and Cloudflare document how to host your own.

MCP on hosting platforms at a glance

Platform	Its own MCP server	Sign-in	What your agent can do	Guide to host your own server
Eurobase	`https://api.eurobase.dev/mcp`	OAuth, with access picked per project	Create and deploy projects, set public variables, publish Functions, read logs	Not documented
Vercel	`https://mcp.vercel.com`	OAuth	Manage projects and deployments, read deployment logs, search docs	Yes, on Vercel Functions
Cloudflare	`https://mcp.cloudflare.com/mcp`	OAuth or an API token	Call the whole Cloudflare API	Yes, on Workers
Render	`https://mcp.render.com/mcp`	OAuth or an API key	Create services, trigger deploys, query databases, read metrics and logs	Not documented
Railway	`https://mcp.railway.com` or a local server	CLI login or OAuth	Create projects, redeploy, deploy templates	Not documented
Netlify	`https://mcp.netlify.com/mcp` or a local server	Not documented	Create, deploy and manage projects	Not documented

Details are from each provider's own documentation, checked on 9 and 10 November 2026.

How MCP works

MCP was released as open source by Anthropic on 25 November 2024. The current revision of the specification is dated 28 July 2026. It names three roles:

Host: the AI application you use, such as Claude Code or VS Code.
Client: the connector inside the host that talks to one server.
Server: a program that offers tools and data to the client.

A server can offer tools, which are functions the model may run, resources, which are data the model can read, and prompts, which are reusable templates. Messages are JSON-RPC 2.0.

There are two standard transports. stdio runs the server as a local process that your AI app starts on your computer, and it reads credentials from the environment. Streamable HTTP sends each message as an HTTP POST to one endpoint on a remote server. A local server usually serves one client, while a remote one serves many. That's why hosting platforms run remote servers: you add a URL, not a program.

For remote servers, the specification builds sign-in on OAuth 2.1. The MCP server acts as a resource server, an authorization server issues the access tokens, and your AI app finds that authorization server through metadata the MCP server publishes. In practice, you add the URL, your browser opens, and you approve access on the platform's own page.

Use 1: your agent manages your deployments

This is the use most developers meet first. You connect your host's MCP server to your coding agent, and the agent can then do what you'd otherwise do in a dashboard. On most platforms that means:

creating a project from a repository,
starting a deployment and watching it finish,
reading build and request logs when something breaks,
setting environment variables,
rolling back to an earlier version.

The value is the closed loop. The agent writes the code, ships it, reads the failure and fixes it, without you copying log lines into the chat.

Platforms draw the line in different places. Eurobase's agent can deploy, redeploy, promote and roll back, but it can't delete anything, and secret values never pass through it: you type them into the project's Secrets yourself. Vercel's docs say plainly that connecting grants the AI the same access as your Vercel user account. Render's docs warn that its server can run destructive operations, including changing a service's environment variables and triggering deploys.

For a side-by-side of every platform's server, read the best cloud MCP servers. For step-by-step setup with one agent, read how to deploy from Claude Code, from Codex or from Cursor.

Use 2: the platform runs your own MCP server

If you're building an MCP server for your own product, it needs a home on the internet like any other web service. Two platforms publish dedicated guides.

Vercel runs MCP servers as Vercel Functions, with a mcp-handler package for Next.js and other frameworks. Each tool call runs inside one function, so the function's maximum duration limits it: 300 seconds by default, and up to 800 seconds on Pro and Enterprise.
Cloudflare runs remote MCP servers on Workers over Streamable HTTP, from a template you can deploy from the dashboard or the CLI. Its current handler is stateless, so data that must outlive a request goes into KV, D1 or a Durable Object.

Eurobase doesn't offer an MCP server template or a guide for hosting one.

What to check before you connect an agent

How access is scoped. Look for a sign-in that lets you pick projects and read or write rights, rather than one that hands over your whole account.
What the agent can't do. A server that can't delete projects or read secrets limits the damage from a bad prompt.
Where secrets go. Values you paste into the chat end up in the model's context. Prefer a host where you add them in the dashboard.
How you revoke access. Removing a server from your agent's config may leave its token valid. Find the place on the platform where you revoke the connection.
Prompt injection. A tool that reads outside content, such as a web page or an issue, can carry hidden instructions. Vercel's docs give the example of an instruction to copy private deployment logs to an outside address.
Where your app ends up. The MCP server only starts the deployment. Your data lives wherever the platform runs.

Where Eurobase fits

Eurobase is a hosting platform from a business registered in Warsaw, Poland, running on European infrastructure operated by Scaleway, a French provider. Its MCP server is remote, at https://api.eurobase.dev/mcp, and Claude Code and Codex are tested end to end. Cursor, Zed and VS Code use the same sign-in but aren't yet tested end to end.

When you sign in, Eurobase opens in your browser and you choose which projects the agent may see, which workspaces it may create projects in, and read or write for each permission. With write access it creates a project from your repository, deploys it, and follows the deployment until the site answers at <name>.eurobase.dev. A request cut off halfway is safe to repeat: the retry carries the same key, so you end up with one project, not two. You can revoke access in Connected apps, and the agent loses it on its next request. A CLI, npx eurobase, offers the same actions from a terminal.

Best for: apps that an agent builds and ships, and that should stay on European infrastructure.
Price: the MCP server works on the Free plan. Starter costs €10 and Pro €20 a month + VAT per workspace. Nothing is billed by usage today. Once a limit is reached, new work is refused, apps on V8 isolates and Functions pause until the monthly reset, and containers and static sites stay online.
What it lacks: preview deployments for the agent to test on, managed Postgres, MySQL or Redis, Dockerfile builds, and a guide for hosting your own MCP server. Its ISO 27001 and SOC 2 work is still under way, so neither is held yet.

Frequently asked questions

What does MCP stand for?

Model Context Protocol. It's an open standard for connecting AI applications to outside tools and data, first released by Anthropic in November 2024 and now supported by Claude, ChatGPT, VS Code, Cursor and many other apps.

Is an MCP server the same as an API?

No. An MCP server usually sits in front of an API. It describes its tools in a standard format that any MCP client can discover, and when the model picks one, the server calls the platform's API and returns the result. Render's docs describe its server this way.

Do I need to install anything to use a hosting platform's MCP server?

Usually not. Remote servers run on the platform, so you add a URL to your agent and sign in through your browser. Railway and Netlify also offer local servers that run on your computer, for setups that need one.

Is it safe to let an AI agent deploy to production?

It can be, if you limit what it can reach. Start with read access, grant write access to one workspace or project, keep secrets out of the chat, and make sure you can roll back. Only connect servers whose address you've checked on the provider's own site.

Can I host my own MCP server in Europe?

Yes, on any platform that runs a web service in a European region. Vercel and Cloudflare publish MCP-specific guides, but both are US companies. If the provider must be a European business, check the company on the terms, not only the region.