SSH MCP Server v2
SSH MCP Server is a security-first Model Context Protocol server that gives LLM agents controlled SSH access to remote hosts — with command classification, policy-based authorization, human-in-the-loop approval, and full audit logging.
The risk this server exists to manage. Giving an LLM shell access on a remote host puts private data, untrusted input and network egress in one place — Simon Willison's "lethal trifecta". Prompt injection has no general fix, so ssh-mcp assumes any command may be attacker-influenced: it classifies before executing, authorizes against a role × host-group matrix, gates destructive work behind approval, and records the decision either way. That narrows the blast radius; it does not remove the risk. Two things stay yours: never point it at a root account, and never set
autoapproval on a production profile. SECURITY.md has the full threat model.
Quick Start
1. Install
npm install -g ssh-mcp
2. Configure
Without a config the server still starts, so a client or directory can complete the MCP
handshake and read tools/list — but every tool call is refused until you configure it,
with a message naming the path below. Nothing runs on a host until this step is done. The
one exception is reload-config: write the file, then call it with action: "reload" (or
restart the server) to configure a server that is already running — see
Reloading.
Create the config file at the path for your platform:
[defaults]
defaultProfile = "dev"
approvalMode = "ask-destructive"
[[profiles]]
name = "dev"
host = "192.168.1.100"
port = 22
user = "deploy" # NOT root!
auth = "key"
keyRef = "~/.ssh/id_ed25519"
role = "admin"
approvalPolicy = "auto" # dev is permissive
chmod 700 ~/.config/ssh-mcp && chmod 600 ~/.config/ssh-mcp/config.toml
The config decides which hosts, roles and policy rules this server honours, so it checks that nobody but you can read it — and treats the two platforms differently, because the question has a much clearer answer on one of them.
Linux and macOS: enforced. The mode check above, on the file and the directory —
which is why chmod 700 is in that command, since mkdir -p under the default umask
leaves the directory 0755. The server refuses to start otherwise. "Only the owner" is
unambiguous here and chmod is a one-line fix.
Windows: split by what the ACL actually allows. There are no mode bits, so the ACL is read instead — and read exposure and write exposure are not treated alike, because Windows is much clearer about one of them than the other.
A config under %APPDATA% inherits access for you, SYSTEM and Administrators and needs
nothing done to it. One created elsewhere does not: a file under C:\ inherits read for
every local account and modify for every authenticated one. The message names the two
icacls commands that fix it either way.
Read exposure is reported rather than refused because that is where Windows is genuinely muddier than POSIX, and refusing over it blocked a config at the documented location (#138). Write exposure is refused because it is not muddy at all: another account being able to rewrite the file that decides which hosts, roles and approval policy this server honours is an authorization bypass, not a disclosure.
Two flags move the whole thing: --strictConfigAcl refuses everything the check objects
to, read-only grants included; --allowUncheckedConfigAcl reports everything and refuses
nothing. Neither combination leaves you without an exit, which is the lesson of #138.
Exit statuses
A supervisor that treats any non-zero status as a failure needs no change. One
that matched on 1 to detect a startup problem should match on 2 as well.
Starting with nothing configured is not an exit-2 condition, as of the release that
added introspection without a config: the server starts so it can be described, and
refuses each tool call instead, except reload-config. A supervisor that used a non-zero
exit to catch an unconfigured deployment should watch for starting unconfigured on
stderr, or read configured from GET /health when running the HTTP transport.
3. Set credentials via environment variables
export SSH_MCP_PASSWORD="your-password" # if using auth=password
# OR use SSH agent (recommended):
export SSH_AUTH_SOCK="$SSH_AUTH_SOCK" # already set if agent running
4. Connect from your MCP client
Claude Code:
claude mcp add --transport stdio ssh-mcp -- ssh-mcp
Claude Desktop / Cursor / Windsurf:
{
"mcpServers": {
"ssh-mcp": {
"command": "ssh-mcp",
"env": {
"SSH_MCP_PASSWORD": "your-password"
}
}
}
}
Never pass passwords as CLI arguments — they're visible via ps aux. Use env vars, config files, SSH agent, or OS keychain.
Tools (15)
Streaming file transfer
sftp-upload/sftp-download move file contents through the model's context:
the text is an argument on the way out and a response on the way back. That is
what you want for a config snippet and exactly what you do not want for a 200 MB
tarball or anything binary.
The two also differ on what they do to an existing destination, and the approved
string now says which is which. sftp-upload replaces unconditionally and spells
--overwrite every time, because that is what it always does; sftp-upload-file
refuses unless you pass overwrite: true, and only then carries the flag. Since
sftp-upload takes its content as an argument rather than naming a local file,
its string also carries --bytes=<n> --sha256=<32 hex> — without that, two
uploads to the same path are the same string, which means one approval covers
both and an auditor cannot tell which set of bytes landed. The digest is 128
bits rather than a short prefix because that claim has to hold against a caller
who picks both payloads, not only against an accidental repeat.
[policy].denylist patterns are tested against that whole string and, for every
SFTP tool, against the remote path on its own — as given, lexically normalized,
and read the way Windows reads it (see "Policy Engine"). A rule written for the
path, such as authorized_keys$, therefore does not depend on where the string
puts it or on how the file is spelled. A rule anchored on the whole string
(^sftp:upload /root/.*$) still does: that layout is ours to change, so anchor
on the path instead.
sftp-upload-file and sftp-download-file stream between the remote host and
local disk instead. Neither the bytes nor a base64 encoding of them ever reaches
the model — the response is a byte count and two paths.
The two transfer tools are off until you configure defaults.transferRoot,
and refuse with an explanation until then. (sftp-list has no local side and
needs none of this.) That directory is the whole of their local reach:
[defaults]
transferRoot = "/srv/ssh-mcp-transfers"
transferMaxBytes = 268435456 # 256MB per transfer
transferTimeoutMs = 300000 # 5min with no bytes moving
Checked on every call, and refused rather than degraded if any of it fails: the
directory must be 0700 and owned by the account running this server, with no
group- or world-writable parent, and must not overlap the ssh-mcp installation,
the config directory, the audit log directory, or ~/.ssh. Caller paths are
confined to it, symlinks are refused rather than followed, and a download is
staged as a sibling .part file and published atomically, so a failed transfer
never leaves a half-written file at the destination.
transferTimeoutMs is an idle budget, not a total one: it bounds one
metadata round-trip, or one stretch of the copy with no bytes moving, and is
re-armed on progress. A slow but live transfer of a large file survives it; a
stalled channel still fails within one window. That is why the byte cap above
does not have to be divided by it — a total budget would have made a 256MB cap
mean "only if the link sustains 900 KB/s".
Not available on Windows, where the transfer root cannot yet be verified private; the two transfer tools refuse there rather than writing into a directory other accounts may be able to read.
sftp-upload-file takes an optional mode (1–511; setuid, setgid and the
sticky bit are refused). Omitted, a new remote file is published 0600; an
overwritten one inherits the replaced file's permission bits — its permission
bits only, not its setuid or setgid.
What a transfer is authorized for includes these arguments: the string the
policy engine classifies, the approval prompt a human reads, and the audit
record all name --overwrite and --mode when they are given. So approving one
upload to a path does not approve a different one to the same path, and an
approval grant (approvalGrantTtlMs) cannot be replayed with a different mode.
Interactive Sessions
Sessions maintain state (CWD, environment variables) between commands:
Agent: open-session(name="deploy", type="interactive")
Agent: run-command(session="deploy", command="cd /opt/myapp")
Agent: run-command(session="deploy", command="git pull") # runs in /opt/myapp
Agent: run-command(session="deploy", command="npm ci") # CWD persists
Agent: close-session(name="deploy")
Background Sessions
Long-running processes (logs, builds):
Agent: open-session(name="logs", type="background", command="tail -f /var/log/syslog")
Agent: read-session-output(name="logs", lines=20) # poll
Agent: close-session(name="logs")
Remote host support
Tested against Linux (Debian/bash, Alpine/busybox ash), Dropbear, and Windows OpenSSH on Windows 11.
Interactive sessions require a POSIX shell (sh, bash, ash, zsh). They work by
bracketing each command with printf markers and reading $? and $PWD from a
trailer — none of which exist in cmd.exe, the default shell for Windows
OpenSSH. Opening one against such a host fails immediately with an explicit
error rather than timing out; everything else works normally.
Setting PowerShell as the OpenSSH DefaultShell does not help: the protocol is
POSIX-specific, not merely non-cmd.
Configuration
Profile options
[defaults]
defaultProfile = "dev"
sessionMaxPerConnection = 5
sessionIdleTimeoutMs = 600000 # 10min
sessionBackgroundMaxMs = 3600000 # 1hr
commandTimeoutMs = 60000
commandMaxChars = 5000 # 0 = unlimited, the config spelling of --maxChars=none
commandMaxOutputBytes = 1048576 # 1MB
connectionIdleReapMs = 900000 # 15min
commandQuotaPerDay = 0 # 0 = unlimited; circuit breaker for runaway agents
approvalGrantTtlMs = 0 # 0 = always prompt; see "Approval Grants"
approvalMode = "ask-destructive" # auto | ask-destructive | ask-all | deny
# transferRoot = "/srv/ssh-mcp-transfers" # enables the streaming file tools; see above
transferMaxBytes = 268435456 # 256MB per streaming transfer
transferTimeoutMs = 300000 # 5min with no bytes moving (idle, not total)
[[profiles]]
name = "prod-web-1"
host = "10.0.1.50"
port = 22
user = "deploy"
auth = "agent" # agent | key | password | keychain
keyRef = "~/.ssh/id_ed25519" # for auth=key
keychainEntry = "ssh-mcp/prod" # for auth=keychain (requires @napi-rs/keyring)
via = "bastion" # ProxyJump — route through bastion profile
group = "prod" # Policy tier: prod | staging | dev, or your own (see [policy])
workdir = "/var/www"
trustedHostKey = "SHA256:..." # Pin host key (optional)
tty = false
role = "operator" # viewer | operator | admin
readOnly = false
approvalPolicy = "ask-all"
cert = false # SSH CA cert auth — auto-detects keyRef-cert.pub
announceAgent = true # Send AI_AGENT=ssh-mcp to this host (see below); clear it for a host you do not control
sessionMaxPerConnection = 3 # per-profile override
sessionIdleTimeoutMs = 300000 # stricter for prod
commandQuotaPerDay = 200 # per-profile override
maxChars = 2000 # per-profile override; stricter for prod
transferMaxBytes = 16777216 # per-profile override; transferRoot is not per-profile
transferTimeoutMs = 60000 # per-profile override
# Optional. Merged over the built-in role matrix; see "Policy Engine" below.
# roleBindings is keyed by role and then by tier, so the block below changes
# operator on prod and leaves operator's other tiers, and viewer and admin,
# on their defaults.
[policy]
denylist = ["^terraform\\s+destroy", "\\.ssh/authorized_keys$"]
[policy.roleBindings.operator]
prod = ["read-only", "safe", "destructive"]
Unknown sections and keys are a startup error, not a warning, so a typo cannot
leave you running defaults you thought you had overridden. That extends to role
and tier names: every one you write under [policy.roleBindings] has to be
reachable by some profile, and every profile's role and tier has to resolve to
real bindings. Both directions are checked at startup and on every reload.
Every profile needs a name of its own. Two profiles with the same name are a startup
error, and a reload refuses them too; the error names the two entries by index
(profiles[2].name repeats profiles[0].name), not by the name. Before this release such a
config started, and the first of the two was used.
Host-side identification
Every command this server runs announces itself to the host as the environment
variable AI_AGENT=ssh-mcp, so an operator can tell an agent's session from a
person's from anything on the host that reads the session environment — a
/etc/profile snippet or a ForceCommand wrapper, for instance. It covers all
three channels a command can run on: one-shot commands, background sessions and
interactive sessions. SFTP transfers do not carry it, because the SSH library
this server uses offers no way to send an environment request on a subsystem
channel.
The variable only lands in the session if the host opts in with
AcceptEnv AI_AGENT in sshd_config. Without that line sshd ignores it and the
session environment is unchanged — measured against Dropbear, which has no
AcceptEnv mechanism at all, and against Windows OpenSSH on a default
configuration: the command runs, the exit code is unchanged, the variable is
simply absent.
The request is nevertheless sent to every host, opted in or not. AcceptEnv
is the server's policy about what it stores; it is not a gate on what this client
transmits. A host that has not opted in — including one that is hostile — can
still see on the wire that an agent rather than a person is driving the session.
That is why announceAgent = false exists on the profile: set it for any host you
do not control and nothing is sent to it. It is on by default, because the
operators who benefit are the ones running the hosts you do control.
No version is sent, deliberately. A version would narrow a hostile host to one specific build; the tool's name alone does not. (The SSH transport already discloses the library version in its identification string, which is one library's version across every program that uses it.)
Treat the variable as a courtesy label, not an attestation. Any SSH client can set it and any agent can omit it, so it is useful for attributing sessions you already trust and must not be used as an authorization or intrusion-detection input. Authoritative attribution belongs to the key or principal that authenticated.
ProxyJump (Bastion)
Reach internal hosts behind a bastion/jump server. The via field specifies a profile name to tunnel through:
[[profiles]]
name = "bastion"
host = "bastion.example.com"
user = "deploy"
auth = "agent"
[[profiles]]
name = "internal-db"
host = "10.0.1.50" # private IP — not directly reachable
user = "dbadmin"
auth = "key"
keyRef = "~/.ssh/db_key"
via = "bastion" # tunnel through bastion
No agent forwarding — only a TCP tunnel via forwardOut. The bastion stays connected and reusable for multiple internal hosts.
SSH CA Certificates
For enterprise setups with a central SSH Certificate Authority:
[[profiles]]
name = "prod-db"
host = "db.internal"
user = "admin"
auth = "key"
keyRef = "~/.ssh/id_ed25519"
cert = true # enable CA cert auth
The certificate file is auto-detected using OpenSSH convention (keyRef + -cert.pub, e.g. ~/.ssh/id_ed25519-cert.pub). You can override the path with SSH_MCP_<NAME>_CERT env var. The cert is concatenated with the private key per ssh2 convention.
Credential Resolution Order
- SSH agent (
SSH_AUTH_SOCK) — no key material in process memory - OS keychain (macOS Keychain / Windows Credential Manager / Linux Secret Service) — requires
auth = "keychain"and@napi-rs/keyring - Environment variables —
SSH_MCP_PASSWORD,SSH_MCP_KEY,SSH_MCP_SUDO_PASSWORD, or profile-specificSSH_MCP_<NAME>_PASSWORD - Key file —
keyRefpath orSSH_MCP_KEYenv var
Never CLI arguments. v2 removes --password, --sudoPassword, --suPassword entirely.
Reloading
The config file is read at startup, and again only when you ask: through the
reload-config tool with action: "reload", or by sending the process SIGHUP
(kill -HUP <pid>) on every platform but Windows. On Windows SIGHUP means the console
is closing, so it does not reload there; use the tool. Nothing watches the file, so a
half-written file never becomes live policy. A reload reads the same path the server
started with: --config, or the default path for your platform.
reload-config takes one argument, action:
"status"(the default) reports the config path, where the live config came from (file,unconfiguredorflags), when it was loaded, the sha256 of the bytes it was parsed from, and the result of the last reload. It never shows the file's contents."reload"re-reads the file, applies it if every check passes, and reports what changed: profiles added, removed and changed (by name and field), whether the policy rules changed, and which connections it closed, why, and with which sessions. The same report goes to stderr.
The tool asks for no approval and spends no command quota. A caller can only make the server re-read what the operator wrote; the file is the operator's consent.
All or nothing. A reload runs the same checks as startup on the new file before anything
changes: the permission check (mode bits on Linux and macOS, the ACL on Windows, with the
--strictConfigAcl or --allowUncheckedConfigAcl the server started with), the TOML parse,
the strict schema, the check that no two profiles share a name, the role and tier
cross-check, and compiling the denylist. A file with no
profiles fails the schema, as at startup, so a reload cannot leave a server unconfigured.
Two more refusals belong to reloading alone:
defaults.transferRootchanged or removed. A transfer in flight holds the old root, so this needs a restart. Adding atransferRootwhere there was none is allowed: with no root there can be no transfer in flight.- The server was started from
--host/--userflags. Swapping the flag-built profile for whatever file sits at the default path would change where the server's trust comes from without the operator choosing it. Restart with a config file instead.
A refused reload changes nothing: the config, the policy, connections, sessions and approval grants stay as they were. The refusal names the error class and the path, and for a schema error the field paths and issue codes; it never quotes a value from the file.
What applies when. Once every check passes, the new config replaces the old in one step, so no tool call sees part of each. For each open connection, by profile name:
A connection that runs through a bastion that is closing closes too, and so does anything
running through it in turn; dependents close before the bastion they run through. A
[defaults] value reaches every profile that does not set its own, so changing it counts as
a change to each of those profiles. defaultProfile, connectionIdleReapMs and
approvalGrantTtlMs apply at once. Command quota counts are kept, so a new
commandQuotaPerDay applies to what has already been spent. Host keys accepted on first use
are kept too, because the connection registry is updated in place rather than rebuilt. A
new connection that was still being opened when the reload landed is discarded only if the
reload closes its profile (removed, a connection field changed, or its bastion closing); that
call fails with "The configuration was reloaded while connecting; retry the call." Any other
is kept, with the profile as the reload left it, so reloading an unchanged file does not fail
other calls' connects.
Approvals. Every applied reload, even of an unchanged file, retires every approval grant
made before it, so the next run of a granted command asks again. An approval answered after
a reload landed is re-checked against the new rules: the command runs only if its connection
is still open and the new rules still allow it or ask for it, in the same command class.
Otherwise it is refused with CONFIG_RELOADED and no grant is recorded; retry the call to
be asked under the new rules. A decision that needed no prompt is re-checked the same way
when a reload lands while it is being made (while the connection is resolved, or during an
OPA round-trip): a command the old rules allowed, or a grant reuse, runs only if its
connection is still open and the new rules allow it outright, in the same class; otherwise
it is refused with CONFIG_RELOADED. The command quota is charged against the profile as
the reload left it.
Already running. A command or transfer that is running when the reload lands finishes
under the rules it started with, unless the reload closes its connection; the next call is
checked under the new ones. Removing a profile and reloading closes its connection, but that
stops less than it sounds: a background session's command is signalled INT, then TERM,
then KILL, while a one-shot command, or one running in an interactive session, is not
signalled at all and may keep running on the host. Stop such a command with
signal-process first, while the profile still reaches the host (see
SECURITY.md).
Never reloaded. Command-line flags keep the values the process started with: the
transport and its HTTP options, --hostKeyMode, the OPA, audit and OpenTelemetry flags, the
ACL flags and --config itself. The flags that stand in for [defaults] (--timeout,
--maxChars, --sessionMax, --sessionTtl, --commandQuota, --approvalGrantTtl) apply
only while no file is loaded, as at startup, so a server that started unconfigured takes its
defaults from the file once a reload loads one.
Audit. Every reload that runs writes one config:reload record, in one of three forms:
The other fields are fixed: command and binary are config:reload, commandClass is
safe, profile and user are -, and mcpRequestId is the tool call's request id, or
sighup for a reload the signal triggered. Each session a reload closes also writes a
session:close … (reload) record, as the close-session tool does. A connection that fails
to close is reported, not rolled back: the new config is already in effect.
Reloads run one at a time: one asked for while another is running waits for it and then
reads the file afresh, and any asked for while that one is still waiting join it and share
its report and its record, so a burst of requests or SIGHUPs costs at most two reloads.
If something fails after the checks have passed (a defect, not a refusal), the report says
the reload failed after the swap began and may be partly applied; check reload-config
status, and restart if it does not show the new config.
SIGHUP no longer stops the server. Before, its default action ended the process
without flushing the audit log. Now, everywhere but Windows, it triggers a reload and the
server keeps running. A stdio server still exits when its client closes stdin. An HTTP
server started from a terminal is no longer stopped when that terminal closes, since the
hangup it sends is now handled as a reload: stop it with SIGTERM or SIGINT, or run it
under a supervisor.
Policy Engine
Roles
Which column applies comes from the profile's group. Set it explicitly —
without it the tier is guessed from the profile name (prod/staging/dev,
local, test, sandbox), and an unrecognised name resolves to prod,
the strictest tier. A production host named web-01 is therefore treated as
production rather than silently getting dev permissions.
Note what this means for sudo: admin has no privileged on prod, so
privileged-command is refused there by design — including on a quick-start
profile, which has no name to infer from and therefore lands on prod. If the
host is not production, say so:
npx ssh-mcp --host=10.0.0.5 --user=deploy --group=dev
[[profiles]]
name = "build-box"
group = "dev"
Configuring the matrix
The table above is the default, not a limit. An optional [policy] section is
merged over it at startup and on every reload, so granting sudo on a host you have honestly
labelled prod is a reviewable line in a config file rather than a relabelling:
[policy.roleBindings.admin]
prod = ["read-only", "safe", "destructive", "privileged"]
The merge is at role and tier depth. That block changes admin on prod and
nothing else: admin on staging and dev keep their defaults, and viewer
and operator are untouched. Roles and tiers the defaults have never heard of
are added rather than rejected, which is what makes a custom group resolve to
real bindings instead of falling back to the strictest tier:
[[profiles]]
name = "build-box"
role = "admin"
group = "tier-1"
[policy.roleBindings.admin]
"tier-1" = ["read-only", "safe", "destructive"]
Extra deny patterns live in the same section, and are applied on top of the never-allowed list rather than replacing it:
[policy]
denylist = ["^terraform\\s+destroy", "\\.ssh/authorized_keys$"]
Each pattern is a regular expression tested against the command string of every
call. For the five SFTP tools it is also tested against the remote path on its
own, three ways: as given; in a normalized reading, where // and /./
collapse, .. resolves against the segment before it, a trailing / is
dropped, and \ is read as /; and in a Windows reading, tested
case-insensitively. The refusal says which matched.
The Windows reading is what Win32 resolves a path to, measured over SFTP on
Windows 11 (build 26200): the case variants, a trailing run of dots and spaces,
and the ::$DATA default-stream spelling all reach the same file, so a rule
that caught one spelling and not another was a bypass, not a distinction. On
top of the normalized reading it strips those spellings, roots a
drive-relative path (C:x/…) at its drive — where it actually resolves
depends on process state this server cannot see, and rooting is the fail-closed
reading — drops \\?\ and \\.\ device prefixes, and keeps a UNC root
(\\server\share\… → //server/share/…). Tested case-insensitively, it also
catches Authorized_Keys with a rule written for authorized_keys; on a POSIX
target those are different files, but erring toward refusal is the direction
this engine takes everywhere else too.
What the readings do not do:
- resolve anything on the target. A relative path stays relative, so a rule
anchored on
/root/does not see.ssh/authorized_keys; symlinks are not followed; and an 8.3 short name (AUTHOR~1) is not expanded — no lexical reading can, it takes the target's own directory listing to know the long name. - test the local path of
sftp-upload-fileorsftp-download-fileon its own. That side is confined bytransferRoot. The local path does end those two tools' command strings, so a pattern anchored on the end of the string can still match it there, as it always could.
So anchor a path rule on its trailing segments: \.ssh/authorized_keys$ matches
the absolute, relative and Windows spellings alike — including the case,
trailing-dot and ::$DATA variants of each. A pattern written for a command is
tested against SFTP paths too — ^rm refuses an sftp-download of a file
called rmlist.txt.
One performance note: patterns are ordinary regular expressions run against caller-supplied path bytes — bounded by the remote-path length limit, the same exposure the command string has always had. Prefer anchored, linear patterns; a catastrophically backtracking one stalls the tool call it matches.
Because role and tier names are free strings, nothing in the merge itself can tell a new custom role from a misspelling of an existing one. A cross-check at startup does, and these all fail there rather than at the point of use:
- a command class outside
read-only | safe | destructive | privileged, so apriviledgedtypo cannot parse into a grant of nothing and then read as a policy decision when a command is refused; - any unrecognised section or key anywhere in the config, so a block the parser does not understand is an error rather than a clean startup with none of the behaviour you configured;
- a role or tier under
[policy.roleBindings]that no profile uses, so[policy.roleBindings.operater]cannot merge in as a fourth role while the profiles you meant to restrict keep running on defaults; - a profile whose
rolehas no bindings, or whose tier has none under that role, so a host cannot end up onread-onlyfor a reason nobody wrote down.
The last one covers the tier you did not set as well as the one you did. A
profile with no group still resolves to one by name, and that inferred tier
has to exist under the profile's role like any other.
A tier with no bindings for a role grants read-only, and never another tier's
classes. There is no fallback between tiers: while the matrix was compiled in,
falling back to prod meant falling back to a role's strictest cell, but a
[policy] block can write that cell now.
An OPA sidecar is not an alternative route to the same grant. OPA is consulted only for commands the local policy already allows, so it can refuse more but never widen. Widening happens here or not at all.
Command Classification
Every command is classified before execution:
- read-only: Allowlisted commands (
ls,cat,grep,df,stat,systemctl status, ...), and only when every argument is provably data under a grammar declared for that binary — see below - safe: Non-destructive mutations (
npm install,git pull, ...) - destructive: mutations that need approval (
rm -rf /tmp/build, ...) - privileged:
sudo,su,doas,pkexec
read-only is granted by argument grammar, not by binary name alone. An allowlisted
binary's argument list is checked word by word against a grammar declared for it
(src/policy/reader-grammar.ts); an option the grammar does not list, an abbreviation of one
it does, a short-option cluster with an unlisted letter, an argument carrying an unquoted
shell glob (*, ?, [), or an operand shaped like something
the binary would write to, all fall the whole command to safe instead. This is what a
readOnly profile is confined to, what viewer holds on prod and staging (on dev the viewer
holds safe too), and what read-command
requires — all of these refuse safe except the viewer on dev, so a command that used to
run under the old, name-only
rule can now be refused. The refusal names the rejected word (for example, `journalctl` is read-only only with the options and operands its grammar lists; `--foo` is not accepted there.), so the caller can see which spelling of the same command still works. See
SECURITY.md for the residual risks this leaves open.
A reader also falls to safe when its command carries a shell metacharacter — ; & |
< > ` $ ( ) { }, a newline or a carriage return — quoted or not, because
the allowlist vouches for the binary, never for what a shell might do with the rest of the
line. The refusal names the first such character. For sftp-list and sftp-download it is
in the remote path, so /srv/backup(old) is refused for a readOnly profile with
The remote path contains `(`, which counts as a shell metacharacter here, so `sftp:list` is read-only only for a path without one.
A separate forbidden list is never allowed, whatever the role or approval
policy: rm -rf /, mkfs, dd of=/dev/, shutdown, curl|sh, fork bombs,
writes naming /etc/cron, /etc/systemd or any user's authorized_keys,
iptables -F, and recursive chmod 777 / / chown /. A write names one of
those paths when it is the remote path of an upload tool, the target of a shell
redirect, or a destination of tee, cp, mv, install or dd of= — in any
spelling the path readings above resolve. The two /etc families match by
prefix, so /etc/crontab is covered along with everything under /etc/cron.d,
and a sibling that merely starts like the directory (say /etc/cronography) is
refused too — the fail-closed direction. Writes spelled any other way (sed -i,
rsync, cmd.exe's copy, move and robocopy, a program handed to an
interpreter, a path the target computes) are not on this list; class and approval
govern them, and the operator denylist can close them. Add your own patterns via
the policy denylist; an invalid pattern fails at
startup rather than degrading silently.
Approval Modes
auto— no prompts (dev only!)ask-destructive— prompt for destructive/privileged (default). Narrower than it sounds: outside the never-allowed list,destructiveis onerm -rf /pathpattern,findwith a write/exec flag, an unresolvable command word, a program handed to an interpreter this server cannot read (python3 -c,perl -e,node -e, a program arriving on a pipe — but notawk, whose program is not read), andsftp-upload/interactiveopen-session— elevation classifiesprivileged, which also prompts. Ordinary writes, service control and signals do not. See SECURITY.md before relying on this in production.ask-all— prompt for every commanddeny— reject destructive/privileged commands outright (no prompt)
Approval Grants (just-in-time)
approvalGrantTtlMs lets one explicit approval cover repeats of the exact
same command on the same profile for a bounded time (e.g. 300000 for five
minutes). It exists because approving rm -rf /tmp/build every few seconds
during an iterative task trains you to click through prompts — which is worse
for safety than a grant you chose deliberately.
A grant is bound to the exact command text, the profile and the command class:
approving rm -rf /tmp/build does not cover rm -rf /tmp/build-prod, the same
command on another host, or the same command escalated to sudo. Runs covered
by a grant appear in the audit log with approver: "jit-grant", so they stay
distinguishable from a fresh human answer.
Off by default (0 = always prompt). Auto-approval weakens the gate that
makes destructive commands safe, so turning it on should be a decision.
Answering the prompt
Approval goes through the MCP elicitation request, so what you see is your client's dialog. Accepting it approves the command — there is no second field to fill in.
You have 10 minutes to answer. Past that the request expires and the command
is refused rather than left pending, and the refusal says so; the prompt may
still be open in your client, in which case run the command again once you are
ready. If your client does not support elicitation at all, every destructive and
privileged command is refused with APPROVAL_UNAVAILABLE naming that cause —
approval fails closed by design.
Command Quota
commandQuotaPerDay bounds how many commands a profile may run in a rolling
24-hour window (0 = unlimited). The approval gate stops destructive commands
and the HTTP rate limiter caps request rate, but neither bounds total work — a
prompt-injected agent looping over allowed commands stays under both. The quota
is the circuit breaker for that case.
Counted after policy allows a command and before it runs, so a denied command does not spend budget. The window slides rather than resetting at midnight, which would let an agent spend a full quota just before the reset and another immediately after.
External Policy Engine (OPA)
For organizations that standardize on Open Policy Agent / Rego:
ssh-mcp --opaUrl=http://localhost:8181
When --opaUrl is set, commands the built-in engine allows are additionally
evaluated by OPA. OPA can only narrow. A command the built-in engine has
already denied returns that denial without OPA being consulted at all, so a
sidecar answering allow cannot grant a class the role bindings withhold. To
widen, edit [policy].
An outage falls back to the local decision and logs one warning per minute. That is the
default because OPA is an additional deny layer and stopping all work would be the worse
failure — but an operator who deployed OPA as the authorization gate loses that gate
during the outage, and the only signal is a stderr line MCP clients usually discard.
--opaFailClosed makes the gate being down mean no; the refusal carries ruleId: "opa-unavailable" so the audit record says the gate was down rather than implying a policy
refused the command.
The request shape is modeled on the AuthZEN Access Evaluation contract:
{
"input": {
"subject": { "role": "operator", "profile": "prod-web-1" },
"action": { "tool": "run-command", "commandClass": "destructive" },
"resource": { "command": "rm -rf /tmp/cache", "binary": "rm", "host": "10.0.1.50" },
"context": { "readOnly": false }
}
}
For the five SFTP tools, resource also carries three path keys: remotePath,
the path as given; and remotePathNormalized and remotePathWindows, the same
two readings the local denylist tests, so a rule can match a canonical spelling
without parsing command or reimplementing either reading in Rego. The two
upload tools add remoteWrite: true, the fact the local policy refuses the
never-allowed write paths on — a rule can gate writes without parsing command
either. Other tools omit all four, and so does an empty path. remotePath is as given —
/root//.ssh/x arrives spelled that way, so a startswith rule on a directory
misses it: match trailing segments, or match the normalized keys.
remotePathWindows keeps the case it was given; lower() both sides in the
policy for case-insensitive rules. Match fields rather than the whole
resource object — its keys vary by tool.
OPA responds with { "result": true/false }. If OPA denies (result: false), the command is blocked even if the built-in engine allows it. If OPA is unreachable, the built-in engine's decision stands by default (fail-open, to avoid locking out access); --opaFailClosed refuses instead. A 200 that carries no boolean result counts as unreachable — that is what OPA answers for an undefined document, so a misnamed package or an unactivated bundle is an outage rather than consent.
Example Rego policy (ssh-mcp.rego):
package ssh.mcp
default allow := false
# Admins pass the OPA gate on dev hosts. The built-in policy still applies on
# top: this widens nothing that the role bindings withhold.
allow if {
input.subject.role == "admin"
startswith(input.subject.profile, "dev")
}
# Deny all destructive commands on prod
deny if {
input.action.commandClass == "destructive"
startswith(input.subject.profile, "prod")
}
Security
Threat Model
See SECURITY.md for the full threat model, vulnerability reporting policy, and deployment checklist.
Supply chain
Releases carry signed attestations, published through Sigstore and recorded in its public transparency log. They live in two different stores, which is what decides how each is verified:
Provenance comes from npm trusted publishing: the release workflow authenticates with a short-lived OIDC token and no stored credential, so there is no long-lived npm token to leak.
npm audit signatures # provenance, against an installed tree
npm pack ssh-mcp # the SBOM attestation is bound to the tarball, so fetch it
gh attestation verify ssh-mcp-*.tgz --repo tufantunc/ssh-mcp --predicate-type https://cyclonedx.org/bom
Both flags on the last command are load-bearing. gh attestation verify defaults to the
SLSA predicate, so without --predicate-type it filters the SBOM out and reports nothing
found — and the provenance it would look for instead is in npm's store, not the GitHub
store --repo queries. Use https://spdx.dev/Document for the SPDX one.
Both SBOMs are also attached to each GitHub release, for reading rather than verifying.
Level 2, not 3. Provenance is signed by the generic GitHub-hosted runner —
builder.id is https://github.com/actions/runner/github-hosted — which the build itself
can influence; Build L3 requires an isolated builder it cannot. Reaching L3 is not
currently compatible with trusted publishing: npm turns on its own provenance whenever that
setting is left at its default, and then ignores any externally generated one. So L3 today
would mean returning to a long-lived npm token — trading the property described above for a
level number.
Safe Defaults
- Non-root user in all examples
- TOFU host key verification (accept on first connect, verify after — within one process; see SECURITY.md)
- RFC 9142 algorithm allow-list (no SHA-1, no CBC, no ssh-rsa)
- exec()-only for elevation (
privileged-commandtakes nosession, so sudo never runs in a persistent shell — fixes PTY leak. Interactive sessions exist for ordinary commands; this bullet is about elevation, not about every path being an exec channel) - Sudo via stdin (not argv — fixes process list leak)
- Sanitizer strips CR/LF/NUL from all metadata
- 3-layer redaction (field → regex → entropy) on audit logs
- No CLI-arg secrets (use env vars, keychain, or config)
- Agent announcement is disclosed, not hidden — every command tells the host
AI_AGENT=ssh-mcp, including hosts that never opted in to storing it; clearannounceAgenton a profile pointing at a host you do not control (see Host-side identification)
Hardening Checklist
- Create dedicated low-privilege service account on target hosts
- Use command-specific
sudoersinstead ofNOPASSWD: ALL - Enable
ask-allapproval for production profiles - Restrict network egress on target hosts
- Use
readOnly = truefor monitoring profiles - Review audit logs regularly
- Run
chmod 700 <config dir> && chmod 600 config.toml(Windows: the ACL under%APPDATA%is already restricted)
Transports
stdio (default)
For local MCP clients (Claude Code, Cursor, Windsurf). No network exposure.
ssh-mcp # reads config from XDG path
ssh-mcp --config=/path/to.toml # custom config path
HTTP (optional)
For remote/web clients behind a reverse proxy with TLS:
ssh-mcp --transport=http --httpPort=3000 --bearerToken=secret
ssh-mcp --transport=http --httpPort=3000 --bearerToken=secret --rateLimit=60
Endpoints: POST / (MCP Streamable HTTP), GET /status, GET /health
GET /health answers {"healthy": true, "configured": <bool>}. It stays 200 either way —
healthy is liveness — while configured is false when no profile is set, which is the
case of a config bind mount that silently did not attach: the server binds the port and
refuses every tool call except reload-config. configured turns true once a reload has
loaded a file. GET /status carries the profile list itself and stays behind the
bearer token.
--rateLimit is a budget per client, keyed the same way as the failed-auth budget below,
and every authenticated request spends from it: POST / and the rest of the MCP route,
GET /status, and an authenticated request to a path that does not exist. GET /health
never does. A wrong token never does either, so unauthenticated traffic cannot spend a
working client's budget. When a client's budget is spent the server returns HTTP 429 with
a JSON-RPC error body, and Retry-After gives the seconds until that client's next request
will be accepted.
Failed authentication is throttled separately, and on by default. --rateLimit never saw a
wrong bearer token, because the auth check answers before the limiter is reached — so
guessing ran at network speed. --authFailureLimit gives each client its own small budget,
spent only on a 401; a correct token never consumes from it, so a working client never
throttles itself. Once an address has spent its budget every request from it waits,
including one with the right token — that is deliberate, since answering the guess would
otherwise tell the caller which token was right. Clients are told apart by socket address,
for this budget and for --rateLimit alike. Behind a reverse proxy that means every
client shares one of each, so set --trustProxy when the proxy is yours — the server
prints a warning the first time it sees X-Forwarded-For without it. --trustProxy takes
the rightmost X-Forwarded-For entry, which is the hop the proxy itself appended;
everything to its left came from the client, so reading the leftmost would let a client
choose its own budget or spend a victim's. That only holds if a proxy actually appended
the entry, so the header is read only when the peer is the proxy — bare --trustProxy
means a loopback peer, which is the deployment above; name a proxy elsewhere with
--trustedProxies. When the header cannot be read as an address, or the peer is not
trusted, the server falls back to the socket address and says so once, so a flag that is
not taking effect is not silent. One trusted hop is assumed. A malformed
--authFailureLimit is refused at startup rather than silently disabling the check; only
0 turns it off.
An IPv6 client is told apart by its /64 rather than its full address, for both budgets. A
/64 is usually the smallest block a subscriber is handed, so hosts that share one — a home
or office network, or customers a hosting provider places on one /64 — share a budget, the
way hosts behind one IPv4 NAT do. An IPv6 address that carries an IPv4 address (IPv4-mapped,
or the NAT64 prefix 64:ff9b::/96) counts as that IPv4 client. A NAT64 or SIIT translator
that uses a prefix of its own is not recognised: every IPv4 client it translates shares that
prefix's /64, and so one budget.
Always terminate TLS at a reverse proxy (Caddy/nginx). The server listens on 127.0.0.1 only.
Docker
# Build
docker build -t ssh-mcp .
# Run (config file + env vars for credentials)
docker run -i \
-v ./config.toml:/home/appuser/.config/ssh-mcp/config.toml:ro \
-e SSH_MCP_PASSWORD=secret \
ssh-mcp
Or with docker-compose:
docker-compose --profile app up
The Docker image runs as non-root UID 65532, with a minimal node:22-slim base.
CLI Flags (v2)
Secrets are never passed as CLI arguments.
Migrating from v1
v2 is a breaking release. Passing a removed flag now fails at startup with the replacement, rather than failing later as a confusing auth error.
Tools
Command results now carry status. In v1 a failed command rejected with
Error (code N). In v2 a non-zero exit comes back as an error result including
the exit code and stderr — so an empty response no longer means "it worked".
Flags
Credentials moved off the command line because CLI arguments are world-readable
via /proc/<pid>/cmdline on Linux (CWE-214). Credentials now resolve through an
SSH agent → OS keychain → env var → key file cascade.
Example
// v1
{ "command": "npx", "args": ["ssh-mcp", "--host=1.2.3.4", "--user=root", "--password=hunter2"] }
// v2 — credentials via env
{
"command": "npx",
"args": ["ssh-mcp", "--host=1.2.3.4", "--user=root"],
"env": { "SSH_MCP_PASSWORD": "hunter2" }
}
For more than one host, move to a TOML config file (see Configuration)
and pass --config <path>; profiles carry per-host roles and approval policy.
Host key verification
v1 did not verify host keys. v2 defaults to trust-on-first-use and records the
key in memory, for the life of the process; a later mismatch in that same
process fails the connection. Nothing is written to disk and ~/.ssh/known_hosts
is not consulted, so a restart accepts afresh — see
SECURITY.md. Pin
explicitly with trustedHostKey in a profile, which is the only control here that
survives a restart — and which no host key mode overrides, so --insecureHostKey
is an opt-out only for hosts you have not pinned (test environments only).
Testing
# Start test SSH servers. --build matters: compose reuses an existing image, and a
# fixture change (such as the alpine target's AcceptEnv line) then never lands,
# which shows up as a product failure in the integration suite.
docker compose --profile test up -d --build
# Run all tests
npm test
# Run only unit tests
npm test -- test/unit/
# Run with coverage
npm run coverage
MCP Inspector
npm run inspect
Contributing
See CONTRIBUTING.md. Please follow the security checklist in all PRs.
Support
If you find SSH MCP Server helpful, consider starring the repository or sponsoring!
Listed on
Also on the official MCP registry as io.github.tufantunc/ssh-mcp.
