Remote MCP vs Local MCP

Remote MCP vs Local MCP

MCP servers come in two varieties based on where they run: Local MCP, which runs on your computer, and Remote MCP, which runs on a server somewhere on the internet. The official MCP architecture guide describes local servers as typically using stdio and remote servers as typically using Streamable HTTP.[1]

Start With a Simple Question: “Where Does the MCP Server Live?”

When you start learning about MCP, a natural question arises:

“Does the MCP server run on my computer, or does it run on Figma’s servers?”

Your answer to that question is exactly the difference between Local MCP and Remote MCP.

The location where an MCP server runs determines how it connects, how secure it is, and how much setup it requires.

Local MCP: Runs on Your Computer

Local MCP is an approach where the MCP server starts up on your local machine.

The AI app (the Host) and the MCP server communicate via standard I/O (stdio) — a mechanism where programs exchange text by reading from and writing to their input and output streams. The official MCP documentation describes stdio as a transport for direct communication between local processes on the same machine.[1]

How Local MCP Works

Your Computer
┌─────────────────────────────────────────┐
│  Claude Desktop (AI Host)               │
│        ↕ stdio (inter-process comm.)    │
│  File System MCP Server                 │
│        ↕                               │
│  📁 Your Local Files                    │
└─────────────────────────────────────────┘
Internet connection: Not required

Concrete Examples of Local MCP Servers

MCP Server	What It Does
File System MCP	Read and write files on your computer
Git MCP	Operate on local Git repositories
SQLite MCP	Query a local database
Browser MCP	Control a locally running browser

Local MCP Characteristics

Item	Details
Where it runs	Your computer
Communication	stdio (standard I/O)
Internet required	No
Security	High (data never leaves your machine)
Setup effort	Moderate to high (installation required)
Getting started	Moderate (may require command-line knowledge)

Remote MCP: Runs on a Server on the Internet

Remote MCP is an approach where the MCP server runs on the service provider’s servers. Nothing needs to be installed on your machine — the AI app communicates with the MCP server over the internet.

The communication uses Streamable HTTP, with Server-Sent Events (SSE) available when streaming is needed. The official MCP documentation describes Streamable HTTP as the transport for remote server communication.[1]

How Remote MCP Works

Your Computer                    The Internet
┌─────────────────┐              ┌──────────────────────┐
│ Claude Desktop  │ ←─HTTP─→     │ Figma MCP Server     │
│ (AI Host)       │              │ (operated by Figma)  │
└─────────────────┘              │   ↕ OAuth auth        │
                                 │  📐 Figma data        │
                                 └──────────────────────┘
Internet connection: Required

Understanding Remote MCP Through Figma Native MCP

In 2025, Figma released a Dev Mode MCP server in beta and later expanded MCP access toward remote AI coding agents and IDEs, according to contemporary reporting.[2] I use Figma here as a concrete example of remote MCP.

What Is Figma Native MCP?

Figma is a UI design tool. Designers use it to create buttons, layouts, and icons.

Previously, if an AI needed to work with Figma data, a developer had to build a custom integration from scratch. Figma’s MCP server introduces a path for MCP clients to access Figma design data and Figma Make code context.[2]

Supported AI apps can access Figma data through an MCP client.

The Setup Flow (Beginner-Friendly)

1. Configure the Figma MCP server connection in a supported AI app
2. Authorize access through Figma-side authentication and permission flow
3. The AI app can then reference Figma data through MCP

In the remote pattern, I do not install a local MCP server process on my computer.

MCP Skills (Capabilities): Concrete Real-World Examples With Figma

The features that an MCP server provides are called Skills — or in MCP terminology, capabilities. There are three types. I will explain each using Figma as a concrete example.

Tools: Operations the AI Can “Execute”

Tools are functions that the AI invokes to do something. These are operations that change state — creating, updating, or deleting things.

Figma-Oriented Examples

Operation Type	What Happens	How You’d Use It
Creation	A new screen or element is created	”Create a draft login screen”
Update	An existing element such as text or color changes	”Align this button color with the brand palette”
Spec generation	Design information is turned into implementation notes	”Summarize the implementation spec for this component”

Everyday analogy: “Tools are like telling a robot to ‘put something on the shelf.’ The physical state of the world changes.”

Important: Because Tools “do something,” a confirmation screen may appear before they execute.

Resources: Data the AI Can “Read”

Resources provide read-only data. Nothing is changed or created.

Figma-Oriented Examples

Resource	What You Get	How You’d Use It
Figma file list	Your projects and files	”What projects do I have?”
Component info	Specs for buttons, cards, and other components	”Show me the button design spec”
Frame structure	Layout and layer hierarchy of a screen	”Describe the structure of the login screen”

Everyday analogy: “Resources are like telling a robot to ‘look at the shelf.’ Nothing moves.”

Important: Resources are “look but don’t touch” — they are safer than Tools.

Prompts: Saving Repeated Tasks as “One-Click Templates”

Prompts are reusable templates for common questions or instructions, stored server-side.

Figma-Oriented Examples

Prompt Name	What It Does
design_review	Automatically runs: “Review this Figma design for accessibility”
generate_component_spec	Automatically runs: “Generate an implementation spec for this component”

Everyday analogy: “Prompts are like saving your usual lunch order as a favorite. One tap and the same instruction runs.”

More Remote MCP Examples Beyond Figma

Beyond Figma, remote MCP servers are increasingly used to connect web services to AI apps. The table below is a conceptual mapping of MCP Tools and Resources to service categories.

Service	Tools Examples	Resources Examples
GitHub	create_issue, merge_pr	Repository structure, file contents
Slack	send_message, create_channel	Channel history, member list
Notion	create_page, update_block	Document content, databases

Side-by-Side Comparison

Comparison	Local MCP	Remote MCP
Where it runs	Your computer	Provider’s server on the internet
Communication	stdio	Streamable HTTP, optionally with SSE
Internet not needed	✅	❌
No installation needed	❌	✅
Good for private data	✅ (never leaves your machine)	⚠️ (sent to the server)
Always up to date	❌ (manual updates)	✅ (provider manages updates)
Representative examples	File system, Git, SQLite	Figma, GitHub, Slack

Which Should I Choose?

Local MCP is a better fit when:

• I want to access files or private data on my computer
• I am working in an environment without internet access
• I do not want sensitive data sent to an external server

Remote MCP is a better fit when:

• I want to access data from a web service like Figma or Slack
• I want a simple setup with minimal configuration
• I want to always have the latest features (the provider handles updates)

I can use both simultaneously. The MCP architecture guide describes Hosts managing connections to multiple MCP servers.[1]

Summary

• Local MCP: Runs on your computer, communicates via stdio. Best for local files, databases, and private resources
• Remote MCP: Runs on a server on the internet, communicates via Streamable HTTP. Best for web services and APIs
• Figma MCP: One representative example for understanding remote MCP
• Tools / Resources / Prompts: The three categories of functionality an MCP server can provide: execute, read, and template[3]

Frequently Asked Questions

Q: Can I use Local MCP and Remote MCP at the same time?

A: Yes. In MCP’s architecture, a Host can manage connections to multiple MCP servers.[1]

Q: Can I build my own remote MCP server?

A: Yes. A remote MCP server can use Streamable HTTP and follow the MCP protocol.[1] Remote servers also require more careful authentication, network, and operations design than local-only servers.

Related Links

• MCP Architecture — Host, Client, Server, and transports
• MCP Capabilities — Tools, Resources, and Prompts

References

1. Model Context Protocol, Architecture overview
2. The Verge, Figma made its design tools more accessible to AI agents, September 23, 2025
3. Model Context Protocol, What is the Model Context Protocol?