Guides · Claude Code + MCP
How to set up an MCP server in Claude Code.
Two minutes, one command. Here is the whole thing up front, then the three details that actually matter (scopes, transports, and how to confirm it connected) and, at the end, the shortcut that skips the config file entirely.
The short version: one claude mcp add, restart the session, run /mcp to check. Everything below is the why.
TL;DR · the whole thing in five lines
- →
One command, then verify. claude mcp add <name> (or a hand-written .mcp.json entry), then restart the session and run /mcp to confirm it connected.
- →
Two choices, both defaulted. Scope (local / project / user) sets who can see the server; transport (stdio for a local program, http for a remote URL) sets how the agent reaches it.
- →
Write once, wire per client. MCP is a standard, so the same server works in Cursor, Copilot, Codex and the rest, but each keeps its own config, so today you set it up again for each.
- →
The file layer is the server worth adding. Interlinked Files answers a whole-machine filename lookup in 85 µs across 4.47M files, about 412,000× faster than VS Code and ~480,000× faster than Windows Search.
- →
One install, 19 clients. A signed installer configures every MCP agent on the box; with the index a Claude Code single-file hunt drops from 6m 57s / 71 calls to 16 ms / 1. Everything local is free, forever: the app, whole-machine search, MCP. Hosting starts at $5.99/mo.
What actually happens
One config entry, a running program, a set of tools.
An MCP server is just a small program that hands an AI agent new abilities (reading files, querying a database, searching the web) through one shared protocol. Setup has three moving parts. You write a config entry once. On every launch, Claude Code connects to the server and reads its menu of tools. Those tools then sit in the agent's toolbelt, callable by the model whenever it decides one fits.
Pick the server and how it talks. A local tool ships as a small program you run on your machine: that is a stdio server. A hosted tool lives behind a URL: that is an http (or sse) server. Most of what you install day to day is local.
Add it. One command does it: claude mcp add <name>. For a local server, put -- before the program and its arguments; for a remote one, pass --transport http and the URL. Prefer a file? Hand-write the same entry into .mcp.json. Either way you pick a scope (below).
Restart the session. Claude Code opens a connection to each configured server the moment it launches and asks one question: what can you do? The server answers with its menu of tools, no work from you.
Verify. Run /mcp inside the session (or claude mcp list from the shell). You will see the server, whether it connected, and the tools it offers. From here you just ask in plain language; the model picks the right tool and calls it.
Prefer to see the file the command writes? A project-scoped server is exactly this block in a .mcp.json at your repo root. Commit it, and everyone who opens the project inherits the server:
The two decisions
Scope and transport, in one screen.
Every server setup makes exactly two choices, and both have sensible defaults. Scope is who can see the server: just you, your whole team, or every project you touch. Transport is how the agent reaches it: a local program, or a remote URL. Keep this as your cheat sheet.
Private to you, in this one project. The right choice while you are trying a server out.
Written to a .mcp.json you commit to the repo: every teammate who opens it gets the same server.
Available to you across every project on the machine. Set it once, it follows you everywhere.
A small executable that runs on your machine and talks over standard input/output. The most common kind, and the private one.
A server reached over HTTP, usually with an Authorization header. Runs somewhere else, on someone else's box.
The streaming HTTP variant some hosted servers still use. Same idea as http, older shape.
Rule of thumb: a tool that runs on your own machine is local + stdio, and that is the private path: nothing crosses the network. A hosted tool is http, and you will usually authenticate it the first time with /mcp inside the session.
One server, many clients
The same server works in Cursor. You just wire it up again.
MCP is a standard, so one server drops into any client that speaks it. The two agents most people reach for both do: Claude Code, Anthropic's terminal agent, and Cursor, the AI-native IDE. The catch is that each keeps its own config in its own place, so today you set the same server up once per client.
Same server. A different config file for every agent that wants it. Now multiply that by Copilot, Codex, Windsurf, Zed, Cline, and the rest of the toolbelt you actually run, and by the day a new one ships. The wiring is the tax.
Skip the config file
One install configures 19 AI clients at once.
Interlinked Files is an MCP server, so everything above applies, but you never open a JSON file. One signed installer detects the AI clients on your machine and writes the config into each of them: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and more: 19 clients wired up in a single step, no per-app editing. And it is not just any server. It is the one an agent reaches for most: the file layer for your whole machine: every repo, every document, even the files a project-scoped search never sees. Local, offline, nothing leaves the disk. Save a file and it is searchable again in about a millisecond.
YOU CAN STILL ADD IT BY HAND. MOST PEOPLE LET THE INSTALLER DO ALL 19.
Measured, not claimed
Why the file server is the one worth wiring.
An agent handed the wrong tools is still slow; an agent handed a fast file layer changes shape. We measured the same single-file task with and without the index, on a real 4.47-million-file machine.
| The searcher | What ran | Time | vs Interlinked |
|---|---|---|---|
| Interlinked Files | one indexed filename lookup | 85 µs | baseline |
| VS Code (Ctrl+P) | same file, same drive | 35 s | 412,000× |
| Windows Search | median of 20 queries | 67 s | ~480,000× |
| ripgrep | same content query, same corpus | 93.8 s | 58,625× |
| Claude Code (no index) | one file, agentic archaeology | 3 m 13 s | 2,276,000× |
Methodology: Ryzen 9 9950X3D · 64 GB DDR5 · NVMe · Windows 11 · 4,470,000 files. Filename lookups are microsecond-scale (85 µs single, 139 µs median of 20 queries). The ripgrep row is the same content query, content-to-content (93.8 s → 16 ms; 58,625× across the five-query set). The Claude Code row is a measured single-file agent run: the agent, given no index, listing directories until it found the file.
Common questions
Setting it up, answered.
What is the difference between a local (stdio) and a remote (http) server?
A stdio server is a small program that runs on your own machine and talks over standard input/output: the common, private kind, since nothing crosses the network. An http server lives behind a URL on someone else's box and usually needs an Authorization header. For anything running locally, reach for stdio.
My server was added but its tools do not work. What do I check?
Run /mcp inside the session (or claude mcp list from the shell) to see whether it actually connected and which tools it advertises. Most failures are a wrong command or arguments in the config, or forgetting to restart the session: Claude Code only reads a server's tool menu on launch.
Do I have to set the same server up again in Cursor and other agents?
Yes. MCP is a standard, so the same server works in Cursor, Copilot, Codex, Windsurf and the rest, but each client keeps its own config in its own place, so today you wire the same mcpServers entry once per client (or let an installer that supports it do all of them at once).
Is any of my data sent to the cloud?
Not for a local stdio server: it runs offline on your machine and nothing leaves the disk. Only http and sse servers reach out over the network, because by definition they live somewhere else.
Why add a file-search server specifically? Is ripgrep not enough?
On a real 4.47-million-file machine the same content query took ripgrep 93.8 s and Interlinked 16 ms (58,625×), and a Claude Code single-file hunt fell from 6m 57s / 71 tool calls to 16 ms / 1. A prebuilt, always-fresh index (re-searchable about a millisecond after you save) is what turns file-finding from the bulk of a session into a rounding error.
Do this today
Set it up once, for every agent.
Install once. One signed installer auto-configures 19 AI clients: Claude Code, Cursor, Copilot, Codex, Windsurf, Zed, Cline and the rest. No claude mcp add, no JSON to edit.
Let it index. Filename search works immediately; content search fills in behind it. At rest the whole engine sits around 44 MB, less than a browser tab.
Ask normally. Your agent picks the right tool by itself. You ask where something is, or what mentions it; the answer comes back in microseconds, whole-machine.
It stays fresh. Save a file and it is searchable in about a millisecond, under 30 in the worst case. The index rides your changes; you never re-index by hand.
Everything (the desktop app and the MCP server every client plugs into) is free forever, on all your devices, no card. Hosting starts at $5.99/mo. Download for Windows.