A local MCP server is a program you install and run on your own computer; your assistant talks to it over stdio, and it only exists on that machine. A hosted (remote) MCP server runs on the vendor’s infrastructure at a URL; your assistant talks to it over HTTP, so there is nothing to install and it works from claude.ai and your phone. Developers often want local. Almost everyone else wants hosted. Here is the plain-English version of why, with a real example of each from the accounting world.
What is a local MCP server?
A local MCP server is a piece of software running on your machine, usually launched by your MCP client as a subprocess. You install it (typically from npm or GitHub), add a block to a config file telling the client how to start it, and supply your own credentials for whatever service it talks to. The client and server then exchange messages over stdio: standard input and output, the same pipes command-line programs have used for decades. Intuit’s QuickBooks Online MCP server is a good example: released in October 2025 as an early open-source preview, it runs locally over stdio, requires your own Intuit Developer app credentials, connects one company at a time, and exposes roughly 144 tools. It is explicitly aimed at developers, and the setup shows it. If the whole concept is new, start with what an MCP server actually is.
What is a hosted (remote) MCP server?
A hosted MCP server is the same idea with the software running on the vendor’s servers instead of yours. Your assistant reaches it over HTTP at a URL, the way your browser reaches a website. There is nothing to install, no config file, and no process to keep alive. LedgerMCP is this shape: a double-entry ledger exposed as 65 MCP tools (25 read, 40 write), listed in Claude’s connector directory. On claude.ai or the mobile app, you open Settings, find LedgerMCP under Connectors, click Connect, and sign in with your email and a 6-digit code; the whole thing takes about a minute. Claude Code, Claude Desktop, and Cursor connect with an API key instead. Our post on how AI agents find MCP servers covers the directory side in more depth.
Stdio vs HTTP: what is actually different?
Stdio and HTTP are the two transports MCP runs over, and the transport decides where the server can live. Stdio requires the client and server to share a machine: the client launches the server and pipes messages to it directly. That makes it simple and fast for local tools, but it also means the server exists only while that machine is on and that client is running. HTTP has no such tie. The server sits at a URL, any MCP client anywhere can call it, and the vendor keeps it running, patched, and updated. When you see “remote MCP server” or “streamable HTTP,” that is this second shape.
How do they compare head to head?
| Dimension | Local (stdio) | Hosted (HTTP) |
|---|---|---|
| Setup | Install software, edit a config file, supply credentials | Connect from a directory or paste an API key, about a minute |
| Where it runs | Your machine, while it is on | The vendor’s servers, always |
| Works from claude.ai / mobile | No | Yes |
| Credentials | Your own API keys in a local config | A sign-in or a scoped key the vendor manages |
| Updates | You pull and reinstall | Automatic, server-side |
| Data location | Stays on your machine (plus the upstream service) | Held by the vendor |
| Example | Intuit’s QuickBooks Online MCP preview | LedgerMCP via the Claude connector directory |
| Best for | Developers, custom and private tooling | Everyone who just wants the tools to work everywhere |
Which is more secure?
Honest answer: they trade different risks. A local server keeps data on hardware you control, which matters for genuinely private tooling. But it also means raw API credentials sitting in a plaintext config file, and it makes you the security team: you patch it, you decide what it can reach, and any process on your machine is one step away from it. A hosted server moves the trust to the vendor, so the real question is what controls the vendor gives you. For LedgerMCP: API keys can be issued read-only, keys are stored SHA-256 hashed, writes support dry-run previews, every write is reversible, and the accounting invariants are enforced in the database itself. Neither shape is automatically safer. Unmaintained local servers and sloppy hosted ones both exist; judge the specific server.
Who should pick which?
Pick local if you are a developer who wants tools against your own code or infrastructure, needs data to stay on-machine, or is building something custom. The QuickBooks preview fits here: if you already have an Intuit Developer account, it is a reasonable way to script against one company file, and our QuickBooks MCP guide walks through the options. Pick hosted if you want the tools available in every conversation, on every device, without babysitting a process: bookkeeping is the canonical case, since you want to ask about your books from your phone as easily as from your desk. The two shapes also coexist fine; many people run local dev tools and hosted services side by side. For a related fork in the accounting world, see MCP servers vs the QuickBooks connector.
Quick answers
What is the difference between a hosted and a local MCP server?
A local MCP server is a program you install and run on your own computer; your assistant talks to it over stdio, and it only works on that machine. A hosted (remote) MCP server runs on the vendor’s infrastructure at a URL; your assistant talks to it over HTTP, so there is nothing to install and it works from claude.ai, mobile, and desktop clients alike.
What does stdio vs HTTP mean for MCP?
They are the two transports MCP runs over. Stdio means the client launches the server as a subprocess on the same machine and pipes messages through standard input and output. HTTP means the client calls the server across the network at a URL. Stdio is inherently local and single-machine; HTTP is how hosted servers work everywhere.
Is a local MCP server more secure than a hosted one?
It keeps data on your machine, which some people prefer, but you also hold raw API credentials in a local config file and you are the security team. With a hosted server you are trusting the vendor’s controls instead, so judge those: LedgerMCP offers read-only keys, SHA-256-hashed key storage, dry-run previews, and reversible writes.
Can I use a local MCP server from claude.ai or my phone?
No. A stdio server only exists on the machine where it is running, so browser and mobile clients cannot reach it. If you want your assistant to have the same tools everywhere you chat with it, you need a hosted server, connected once through the connector directory or an API key.
