# Getting started
`modelcontextprotocol` is a single Rust binary that provides four Model
Context Protocol (MCP) servers over stdio:
| `filesystem` | `mcp-filesystem` | Secure read/write access restricted to allowed directories |
| `fetch` | `mcp-fetch` | Fetch URLs, convert HTML to markdown, robots.txt enforcement |
| `memory` | `mcp-memory` | Persistent knowledge-graph memory as JSONL |
| `shell` | `mcp-shell` | Execute local programs directly with a restricted working directory |
All four implement the MCP `2026-07-28` specification (see
[Protocol](/protocol)).
## Requirements
- A recent Rust toolchain (the crate declares MSRV 1.97 and edition 2024).
- A 64-bit desktop OS; the crate builds on Linux, macOS, and Windows (CI
runs the test suite on all three).
- Node.js 24 is only needed to build and preview this documentation site,
not to use the servers.
## Installation
From crates.io:
```bash
cargo install modelcontextprotocol
```
Or install directly from the source repository:
```bash
cargo install --git https://github.com/maxylev/modelcontextprotocol
```
Both install the `modelcontextprotocol` binary into your Cargo bin
directory. The release profile is tuned for a small binary (about 5 MB).
Verify the install:
```bash
modelcontextprotocol --version
modelcontextprotocol --help
```
## Configure your MCP client
Each server is selected on the command line. Both a subcommand form and an
equivalent flag form are supported, so the binary fits any client's
configuration style. The subcommand form is recommended.
Subcommand form:
```json
{
"mcpServers": {
"filesystem": {
"command": "modelcontextprotocol",
"args": ["filesystem", "${HOME}/Developer"]
},
"fetch": {
"command": "modelcontextprotocol",
"args": ["fetch"]
},
"memory": {
"command": "modelcontextprotocol",
"args": ["memory"]
},
"shell": {
"command": "modelcontextprotocol",
"args": ["shell", "${HOME}/Developer/my-project"]
}
}
}
```
Flag form:
```json
{
"mcpServers": {
"filesystem": {
"command": "modelcontextprotocol",
"args": ["--filesystem", "${HOME}/Developer"]
},
"fetch": {
"command": "modelcontextprotocol",
"args": ["--fetch"]
},
"memory": {
"command": "modelcontextprotocol",
"args": ["--memory"]
},
"shell": {
"command": "modelcontextprotocol",
"args": ["--shell", "${HOME}/Developer/my-project"]
}
}
}
```
See [Command line (CLI)](/cli) for the complete reference, including
options, conflict rules, and environment variables.
## Validate a running server
The fastest offline check is to run the binary without arguments: it prints
the usage summary to stderr and exits with code 1.
```bash
modelcontextprotocol
# modelcontextprotocol: exactly one server must be selected, ...
```
To inspect the live server (tools, prompts, resources), connect any MCP
client. A practical check is to launch the binary and use a generic stdio
client, or exercise it interactively with the MCP Inspector — see
[Verification](/verification) for worked examples.
## Read before you trust
- The [shell server](/servers/shell) executes arbitrary local programs with
the OS permissions of the MCP server process. Its working-directory
restriction is **not a sandbox** and there is **no command filter**.
- The [fetch server](/servers/fetch) can reach **local and internal network
addresses**; there is no SSRF protection.
- Only connect these servers to clients you trust, and never expose them
over an untrusted network. See [Security model](/security).
## Next steps
- [Adding to MCP clients](/clients) — exact configs for opencode, Claude,
Codex, Pi, Gemini, Cursor, and more.
- [Servers reference](/servers/filesystem) — every tool, parameter, and bound.
- [Protocol](/protocol) — what the binary implements and what is verified.
- [OpenRouter E2E](/openrouter-e2e) — the gated real-network acceptance suite.
- [Coverage matrix](/coverage) — the case catalog behind the tests.