fast-mcp-ssh

SSH, SFTP and persistent shells for AI agents. One Rust binary, no runtime.

An MCP server that gives a model real SSH access:

• One connection per host, held open across calls.
• A PTY shell that remembers cd and export.
• SFTP instead of cat > file, and host-to-host copies that skip your disk.
• A screenshot of the remote desktop.
• Regex guards before anything leaves your machine.
• An append-only audit log of every call.

Answers come back as TOON, roughly 40 percent fewer
tokens than JSON on tabular data.

Install

cargo install fast-mcp-ssh

Or a prebuilt binary from the
latest release,
checked against SHA256SUMS.txt. Linux and macOS ship x86_64 and aarch64,
Windows x86_64.

Copy hosts.example.toml to ~/.fast-mcp-ssh/hosts.toml
and fill in your hosts. Keys go in ~/.fast-mcp-ssh/keys/<name>; auth is
key, agent or password.

Wire it up

.mcp.json, or claude_desktop_config.json for Claude Desktop:

{
  "mcpServers": {
    "ssh": {
      "type": "stdio",
      "command": "fast-mcp-ssh"
    }
  }
}

Same block in Claude Code, Cursor, Windsurf, Zed, VS Code Copilot and anything
else that speaks MCP over stdio.

In the MCP registry it is
mcp-name: io.github.klNuno/fast-mcp-ssh.

Tools

host is optional on every tool once [defaults] default_host is set.

Every tool carries MCP annotations (readOnlyHint, destructiveHint,
idempotentHint, openWorldHint) so a client can gate destructive calls.

Host-to-host copy

cp moves a file from one configured host to another. The bytes never land on
your disk and never reach the model, and a sha256 is compared on both ends
before success. Guards cover the destination too, so a read-only target still
refuses the write.

Remote screenshots

shot hands the model an image instead of a wall of text. It uses whichever of
grim, gnome-screenshot, spectacle, ImageMagick import or scrot the
host has, covering X11 and wlroots Wayland, and downscales locally so a 4K
screen is not a multi-megabyte payload.

Protocol

Speaks stateless MCP (2026-07-28) and every revision back to 2024-11-05,
picked per peer. Stateless changes three things:

• Confirmations come back as an input_required result the client answers and
  retries (SEP-2322), because a server may no longer open a request of its own.
  Older peers keep elicitation/create.
• Long calls hand back a task handle to poll (SEP-2663): exec past its 60s
  timeout, tail with follow=true. Clients without the extension keep the
  blocking call.
• tools/list is sorted, so it is byte-identical between restarts, and carries
  a one hour ttlMs (SEP-2549). A client's prompt cache keeps hitting.

Shells are unaffected. A PTY has always been addressed by the host and
session arguments of the call, which is the explicit handle stateless wants.

Security

Guards run before any SSH packet leaves. deny_patterns refuse outright,
confirm_patterns ask the user, a client that cannot answer is denied, and
read_only = true blocks anything that looks like a write.

Paths are checked on both sides: remote reads of keys, shadow files and cloud
credentials, local writes into your ~/.bashrc or an autostart folder. Every
path-taking tool runs both checks, tail included, and re-checks once the
server has resolved the path, so a symlink cannot launder a blocked target. A
path that will not resolve refuses the call.

Host keys are pinned, TOFU by default, strict and per-host fingerprints
available. Every call lands in ~/.fast-mcp-ssh/audit.log as NDJSON, with
credentials scrubbed.

Guards stop accidents, not an adversary who controls the model. Scope the
remote account accordingly. Threat model: SECURITY.md.
Version history: CHANGELOG.md.

Benchmark

50 iterations per scenario, same Linux host, same LAN, same SSH key, client on
Windows 11. Medians, lower is better, measured on 0.5.0. Reproduce with
benchmark/, raw runs in
benchmark/results/.

Both alternatives are Node, so ~250 ms of their cold start is the runtime
booting. The steady-state gap is the connection: fast-mcp-ssh holds one SSH
session per host and opens a channel per call, the other two reconnect. Writes
go over SFTP here, through a cat > file heredoc there.

[^1]: mcp-ssh-manager truncates that response to 12 KB, so it is not
returning the same output. ssh-mcp-server returns raw stdout with no exit
code, which is why its replies are the shortest and why a failed command looks
like a successful one.

Development

cargo install --path .        # build and install from a clone
cargo test                    # unit tests
cargo clippy --all-targets    # no warnings allowed in CI
./scripts/test-sh.ps1         # end-to-end against a real host (Windows)

Never write to stdout outside the MCP transport: a stray println! corrupts
the JSON-RPC stream and the client disconnects without an error. tracing
macros go to stderr and are safe.

License

MIT.