@yawlabs/ssh-mcp
[Add to Yaw MCP]
One click adds this to your local Yaw MCP config so it's available in every Yaw Terminal session. Tested with Yaw MCP 1.1.13 or newer. Or install manually below.
[npm version]
[License: MIT]
Make SSH work for AI tools. MCP server that manages your SSH environment, diagnoses what's broken, fixes it, and gives your agent remote access to anything.
Built and maintained by Yaw Labs.
The problem
AI CLI tools run in subprocesses where SSH is constantly broken. The agent tries to git pull and gets Permission denied (publickey). It tries to SSH into a server and the agent socket is stale. It tries to deploy and the host key changed because the instance was recreated. Every time, the AI has no idea what's wrong and spirals.
This happens across every situation that needs SSH keys:
- Git — clone, pull, push, fetch, submodules, LFS
- Package managers —
npm install, pip install, go get, cargo, composer from private repos
- Server access — SSH, SCP, SFTP, rsync
- Tunneling — port forwarding to databases, SOCKS proxies
- Deployment — Ansible, Terraform, Capistrano, deploy scripts
- Cloud — AWS EC2, GCP, Azure, DigitalOcean, any VPS
ssh-mcp fixes this. It manages the SSH agent, loads keys, diagnoses failures with actionable fix commands, and provides remote operations — all as MCP tools your AI agent can call.
Quick start
Add to your MCP client config:
On Windows wrap with cmd /c since Node 20+ can't spawn .cmd files directly:
The @latest tag makes npx re-resolve against the registry on every spawn, so each MCP session uses the newest published version. Or install globally if you'd rather pin (no auto-update):
Tools
SSH environment management
Tools that fix your local SSH setup so everything else — git, deploys, tunnels — stops breaking.
Diagnostics
Remote operations
Higher-level operations
Tools that wrap common patterns agents build with ssh_exec — faster and less error-prone.
Auto-diagnostics
When any remote operation fails, ssh-mcp automatically runs diagnostics and includes the results in the error response. Your agent doesn't need to call ssh_diagnose separately — it gets told what's wrong and how to fix it right in the error message.
Connection pooling
Remote operations reuse SSH connections automatically. When your agent makes multiple calls to the same host, the first call opens a connection and subsequent calls reuse it. Connections are kept alive for 60 seconds after the last use, then closed automatically.
The pool caps at 100 connections by default, and a connection still being dialed counts against the cap. When the cap is reached the pool closes an idle connection to make room. If every slot is in use or dialing, a tool call waits for one instead of failing:
- A single-host remote tool (
ssh_exec, the SFTP tools, ssh_find, ssh_tail, ssh_service_status) waits up to its timeout (30s for the SFTP tools, which have no timeout parameter) and is next in line for the first slot that frees, so an ssh_exec arriving in the middle of a wide ssh_multi_exec runs between two of its hosts rather than being refused for the whole run. It reports Connection pool is full only if it never won a slot in that time.
ssh_multi_exec runs at most SSH_MCP_MAX_POOL_SIZE hosts at once (default 100) and works through a longer list as slots free up, so raising the cap raises its parallelism; timeout is per host, not a bound on the whole call. When the pool is full it waits, and gives up only when none of its own hosts holds a slot and a full timeout has passed with none of them starting or finishing. The hosts waiting at that point (up to one per parallel slot) and every host still queued then report Connection pool is full; the queued ones are never attempted.
A call reaches Connection pool is full only when other calls — a long ssh_exec, a concurrent ssh_multi_exec — kept every slot from it for a full timeout; a fan-out wider than the cap that has the pool to itself never does, however long it runs. If you see it, rerun once those calls finish (for ssh_multi_exec, the hosts that reported it), or raise the cap with SSH_MCP_MAX_POOL_SIZE=<n> in the server's environment. The error names the variable.
The library's acquire() / withConnection() do not wait by default: they reject at once with a PoolFullError (code: "ERR_SSH_MCP_POOL_FULL", isPoolFullError()) unless you pass { waitForCapacityMs } — see Programmatic usage.
SSH config support
All connections respect your ~/.ssh/config. Host aliases, custom ports, usernames, identity files, and ProxyJump settings are used automatically. If you have Host myserver configured in your SSH config, just pass host: "myserver" — ssh-mcp resolves everything.
ProxyJump / bastion hosts are supported automatically. If your SSH config has ProxyJump bastion for a host, ssh-mcp connects through the bastion transparently. Chained proxies work too.
Host key verification
All remote operations verify the server's host key against ~/.ssh/known_hosts:
- Known host, key matches — accept.
- Known host, key changed — reject (MITM protection). The rejection message distinguishes a genuine key mismatch from "the server offered a key type your
known_hosts entry doesn't cover", so a missing ed25519 line doesn't read as an attack.
- Unknown host — accept, unless strict mode is on (
SSH_MCP_STRICT_HOSTKEYS=1).
That last branch is trust-always, not TOFU. Real trust-on-first-use pins the key it saw the first time and rejects a change afterwards. The connection path never writes to known_hosts — no tool adds an entry as a side effect of connecting — so connecting pins nothing: every connection to a host absent from known_hosts is a "first" use and is accepted, including one where an attacker swapped the key since your last call. Only hosts put into known_hosts out of band get mismatch protection — by you, by ssh-keyscan, or by ssh_known_hosts_fix, the one tool here that does write the file. Call it explicitly to add an entry so future changes are caught.
For stricter environments, set SSH_MCP_STRICT_HOSTKEYS=1 (or true, case-insensitive) to reject unknown hosts. Add them explicitly with ssh_known_hosts_fix first. The older name SSH_MCP_STRICT_HOST_KEY still works and accepts the same values, but prefer the new one: credential scanners such as Yaw MCP's doctor treat any name with a _KEY segment as a secret, and flag the old name as a plaintext credential.
The diagnostic tools (ssh_test, ssh_diagnose) use StrictHostKeyChecking=no for their probe commands. Those probes only run echo SSH_OK — no credentials or data pass through — so the relaxed setting is safe for connectivity testing. Real operations always go through the hostVerifier.
Command policy
ssh_exec and ssh_multi_exec accept free-form shell commands from the agent. For security-conscious deployments, you can restrict which commands run via two env vars, each accepting a comma-separated list of regex patterns:
SSH_MCP_COMMAND_WHITELIST — if set, the command must match at least one pattern, else it's blocked.
SSH_MCP_COMMAND_BLACKLIST — if set, the command must not match any pattern, else it's blocked.
When both are set, the command must pass both checks (whitelist first, then blacklist). When neither is set (the default), all commands are allowed.
This is a guard rail, not a security boundary. A regex over a string that a remote shell will interpret can catch an agent going off-script; it cannot contain a hostile one. It covers only ssh_exec and ssh_multi_exec: the SFTP tools are not gated at all (see Scope). An agent that also has a shell and the credential can bypass any control in this server. If you need a real boundary, enforce it on the remote host: a restricted account, a forced command in authorized_keys, sudo rules, or filesystem permissions.
Patterns are JavaScript regexes. Use ^ and $ for anchored matches; otherwise patterns are treated as substring matches. Commas separate patterns, except a comma inside a character class ([,]), an escaped comma (\,), or one in a {m,n} quantifier, which all stay part of the pattern. Any other literal comma, including one inside a group, must be written \,, [,] or \x2c. In a list of more than one pattern, write a literal comma as \, ([a\,b]) rather than [,]: an unescaped comma in a class is refused whenever another entry follows, so ^foo[,]bar$ works alone but must be ^foo[\,]bar$ (or kept as its own single-pattern variable) beside ^df$.
A malformed entry blocks everything. If any entry in either variable fails to compile, every ssh_exec and ssh_multi_exec call is refused with an error naming the entry, until you fix or unset the variable. The same applies to a character class that may have swallowed a separator: a class holding an unescaped comma is refused when its content suggests a swallowed separator (an unescaped [ before the comma, or a letter, digit, _, ^, [ or another comma after it) or when another entry follows the class, because ^rm[,^reboot,^mkfs[.] and ^rm[abc,>],^df would otherwise compile as one pattern and stop blocking reboot and rm. Escape the stray bracket as \[, or write a literal comma as \, ([a\,b]). This catches likely typos and is not a proof: ^rm[,>] still reads as one class.
While a whitelist is active, a matching command is still refused if it:
- contains shell control syntax anywhere, even inside quotes:
; & | a backtick $( ${ $[ @{ ( ) < > or a line break. Quotes are not interpreted because the remote login shell is unknown and shells disagree about them. So grep 'a|b' file is refused too, and so is git show HEAD@{1}, because @{ runs code in PowerShell.
- starts with a variable assignment, such as
LD_PRELOAD=/tmp/x.so ls or PATH=/tmp ls. This includes the prefix built from the env parameter.
Set SSH_MCP_COMMAND_ALLOW_SHELL_SYNTAX=1 (or true) to turn both checks off and go back to pattern-only matching. Each pattern is then responsible for the entire command, so anchor it at both ends as in the example above. Any value other than 1, true, 0, false or empty is rejected as a misconfiguration.
Choosing what to allow. Only allow programs that cannot run other programs or write files with the arguments you permit. These look read-only and are not:
find runs commands and deletes files (-exec, -execdir, -ok, -delete, -fprint).
awk runs commands (system(), | getline) and writes files.
tar runs commands (--checkpoint-action=exec=..., --to-command, --use-compress-program).
sed (e command, w), xargs, env, sudo, ssh, git (-c, aliases, hooks), less and vim can also launch programs.
cat, tail and head run nothing, but read any file the account can read, including keys and credentials.
A whitelist only sees the command text, so a broad pattern such as ^find allows all of the above.
Denylists are best effort. The same command has endless spellings: ^rm misses /bin/rm, \rm, r''m, command rm, busybox rm, and anything run through sh -c or a script. Pipes to curl and similar exfiltration need no denied word at all. Use a denylist to catch mistakes, never as the control that keeps something from happening.
Blocked commands surface as a clear error mentioning which pattern (or which env var) rejected the call, so the agent can adapt rather than guess. Policy is enforced before the SSH connection opens — no remote process is started for a blocked command.
Scope: policy covers ssh_exec and ssh_multi_exec only
Every other tool runs unchecked. That splits into two very different cases.
The structured read tools are all exempt, but for two different reasons — they don't reach the remote the same way.
ssh_find, ssh_tail and ssh_service_status do build a shell command (find, tail, systemctl), but from typed parameters with every interpolated value shell-quoted, never from free-form agent input. They're exempt for ergonomics: a tight ^ls whitelist would otherwise force you to allow ^find , ^tail , ^systemctl just to keep those tools working — defeating the point of a tight whitelist.
ssh_ls, ssh_stat, ssh_read_file and ssh_download build no command at all. They're pure SFTP (readdir, stat, readFile, fastGet), so — exactly like the mutating SFTP tools below — there is no command string for a regex to match, and these env vars could not gate them even if you wanted them to. They're grouped with the reads rather than flagged as a gap because they don't mutate remote state. One caveat: ssh_download is non-mutating on the remote only — it writes to whatever local path it's handed, and these env vars don't constrain that either.
The SFTP tools that mutate remote state are also unchecked, and that is a genuine gap — not an ergonomics call. ssh_write_file, ssh_upload, ssh_mkdir, and ssh_delete never build a shell command string, so a command-shaped regex has nothing to match. Concretely: SSH_MCP_COMMAND_BLACKLIST="^rm " does not stop ssh_delete, and SSH_MCP_COMMAND_WHITELIST="^ls " does not stop ssh_write_file. Closing this would need a separate path-policy mechanism, which this server deliberately does not have. If you must prevent remote mutation, drop those four tools from your MCP client's tool allowlist (or run a client that gates them) — these two env vars cannot do it.
Policy interaction with the env parameter (ssh_exec, ssh_multi_exec)
When ssh_exec or ssh_multi_exec is called with env: { KEY: "value" }, the values are injected as a KEY='value' ... shell prefix before the command (see the tool descriptions). Policy is checked against the full prefixed command, not the bare command argument — once, before fan-out, in the ssh_multi_exec case. That's the safer ordering at the protocol layer, and it has two effects under a whitelist.
env is refused by default while a whitelist is active. A variable such as LD_PRELOAD or PATH changes what an allowed command actually runs, so the prefix counts as a leading assignment (see above). Remove env from the call. A ^ls pattern will not match the prefixed string in any case, and the rejection explains why.
If you must allow a specific variable, set SSH_MCP_COMMAND_ALLOW_SHELL_SYNTAX=1 and write a pattern that names the variable and spells out the whole command. Never use a generic variable-name pattern, which would admit LD_PRELOAD and PATH:
(That is the double-quoted shell form. The regex itself is ^FOO='(?:[^']|'\\'')*' ls( [A-Za-z0-9_./:=@%+-]+)*$, which is what belongs in a JSON env block after JSON escaping.)
Avoid substring-match patterns like ls. An agent could pass env: { ATTACK: " ls " } to make the final command ATTACK=' ls ' some-other-command, which contains ls. The assignment check refuses that by default, but some-other-command ls would still match, so anchor every pattern.
Blacklists see the prefix too. ^rm blocks a bare rm call, but doesn't block FOO='bar' rm. An env-tolerant anchor closes that one spelling, and like every denylist it stays best effort:
Windows support
On Windows, ssh-mcp uses the OpenSSH Authentication Agent's \\.\pipe\openssh-ssh-agent named pipe automatically when SSH_AUTH_SOCK is not set. No SSH_AUTH_SOCK needed — just make sure the OpenSSH agent service is running.
On the oam runtime, the launcher serves only on oam 0.21.1 or newer. Agent auth needs a net that can dial a named pipe or a Unix socket (that is how ssh2 reaches the agent), which older oam releases could not do. On oam 0.20.2 (Windows) the MCP handshake, ssh_config_lookup, ssh_key_list and ssh_agent_ensure were verified through the launcher, and ssh2's own agent client was verified dialing a named pipe and completing an identities request (against a stand-in agent on a pipe). An agent-only ssh_exec end to end has so far been verified on Node only. Started from Node, the launcher prefers oam (auto); to rule the runtime out when agent auth fails there, set SSH_MCP_RUNTIME=node. Started by oam with no Node on PATH, it can still serve on oam: see Runtime selection. ed25519 and ECDSA keys and ed25519-only servers used to be broken on oam 0.20.2; that is fixed as of 0.21.1 — see oam key support.
ssh_agent_ensure and ssh_diagnose probe that pipe and tell you if the service is down. Remote operations do not: they assume the pipe and let the connection fail on its own if the agent isn't there. That is why a stopped agent service shows up as an auth failure rather than an "agent not running" error until you run the diagnostic tools.
Authentication
All remote operations accept connection parameters:
Auth resolution. An explicit credential wins outright and nothing else is offered. With neither given, ssh-mcp offers the ssh-agent and one on-disk key together — the way the OpenSSH client does — and lets the server pick during the auth exchange. It is not a strict first-match chain past step 2.
- Explicit
privateKeyPath — used alone. The agent is not offered and no other key is read.
- Explicit
password — used alone, same as above.
- Neither given — both of the following are configured on the same connection:
- ssh-agent —
SSH_AUTH_SOCK, or on Windows the \\.\pipe\openssh-ssh-agent named pipe. The Windows pipe is assumed unconditionally; ssh-mcp does not check that the OpenSSH Authentication Agent service is actually running.
- One on-disk key — the first readable path in
ssh -G <host>'s identityfile list. OpenSSH emits that list for every host, including one with no IdentityFile line (it defaults to ~/.ssh/id_rsa, id_ecdsa, id_ecdsa_sk, id_ed25519, id_ed25519_sk), so this is the normal path — not a path reserved for hosts you configured an identity for. ssh-mcp's own built-in list (~/.ssh/id_ed25519, id_rsa, id_ecdsa) is a fallback used only when ssh -G cannot run at all, e.g. no SSH client installed.
Two things worth knowing about step 3:
- Only one on-disk key is ever offered — the first candidate that exists. ssh-mcp does not walk the whole identity list the way
ssh does, so if the first readable key is the wrong one, authentication rests on the agent's keys. Pass privateKeyPath to force a specific key.
- When an agent is configured, an encrypted on-disk key is skipped and the scan moves to the next candidate. The underlying ssh2 library parses
privateKey eagerly and errors with "no passphrase given" on an encrypted key, which would break the common setup of an encrypted key on disk with its decrypted copy loaded in the agent. With no agent configured, the first existing key is loaded regardless of encryption and ssh2 surfaces the passphrase error itself. Because the Windows pipe above is assumed unconditionally, the skip is always in force on Windows.
Example workflows
Agent can't git pull
Host key changed after instance recreation
First-time connection to a new server
Programmatic usage
Runtime selection
The ssh-mcp command — what npx @yawlabs/ssh-mcp runs — is a small launcher. The default runtime, auto, prefers the newest usable oam and falls back to Node when there is none. The minimum is the latest oam release, currently 0.21.1, and the launcher never serves on an older oam.
The default prefers oam again as of oam 0.21.1, which closed the node:crypto gaps it used to trip over: an ed25519 key file, an ECDSA key file, an ed25519-only server and curve25519 key exchange all work there. It was Node-only for one release line (0.18.4-0.18.6), while oam 0.20.2 was the floor; those four gaps are tracked in YawLabs/oam#489, now closed.
Two environment variables control the choice; set them in your MCP client's env block:
SSH_MCP_RUNTIME — which runtime to use. Case-insensitive, surrounding whitespace ignored; any value other than oam or node behaves as auto.
auto (default) — prefer oam. On a Node host the launcher discovers the newest usable oam and spawns it, falling back to Node in-process when none is found. Started by oam at or above the floor it serves in-process, with no second oam. Started by oam below the floor it hands off to Node on PATH (or to a newer oam), or exits with status 1 when there is neither; see the list below.
oam — require oam. Fails loudly: if no usable oam is found, the launcher prints what it found and why each one was passed over to stderr, and exits with status 1 instead of starting the server.
node — always Node. The launcher does not look for oam, and OAM_BIN is ignored.
OAM_BIN — with SSH_MCP_RUNTIME=oam, path to an oam binary to use in preference to discovery, when it is 0.21.1 or newer. If it does not exist, is older, will not run, or does not answer --version with oam's own oam x.y.z banner (another program's version, such as Node's v22.x.y, does not count), the launcher says so on stderr and carries on with discovery.
Discovery (under SSH_MCP_RUNTIME=oam) looks in $OAM_INSTALL_DIR (when set — the oam installer's install target), then %LOCALAPPDATA%\oam\bin (Windows only), then ~/.oam/bin, then every directory on PATH; asks every oam binary it finds for its version, counting only an answer that carries oam's own oam x.y.z banner; and uses the newest one at 0.21.1 or newer. On a tie the one found first wins, so an installed copy beats one on PATH. On Windows only oam.exe counts: an oam.cmd/oam.bat shim on PATH is never run, and is named on stderr when no usable oam is found.
With SSH_MCP_RUNTIME=oam an unusable oam is an error: the launcher names on stderr each oam (and OAM_BIN) it passed over and why — older than 0.21.1, not runnable, or a shim — or the chosen oam that failed to start, and exits with status 1. It does not fall back to Node.
When the launcher is itself started by oam (oam run <path>/bin/ssh-mcp.mjs, which is how Yaw MCP starts an npx @yawlabs/ssh-mcp entry when a recent oam is installed):
- In
auto mode an oam 0.21.1 or newer serves the server itself, in the same process, with no note about unusable key types -- the gap that used to earn one is closed. An oam below the floor never serves there: it hands off to Node on PATH, or to the newest usable oam it can find, or exits with status 1 when there is neither.
- With
SSH_MCP_RUNTIME=oam on oam 0.21.1 or newer, it runs the server on that oam, in the same process. Nothing is discovered or spawned, and OAM_BIN is not read: that oam is the one required. On an older oam it never runs the server there; it hands off to the newest usable oam it can find, or exits with status 1.
- With
SSH_MCP_RUNTIME=node it hands the server off to Node on PATH, whatever the oam version, and exits with status 1 when there is none.
Whenever an oam host starts a child (the handoffs above), the launcher removes --permission and --allow-* flags from the child's inherited NODE_OPTIONS. oam (since 0.18.0) adds its own permission flags to every child's NODE_OPTIONS, and Node refuses some of them there (--allow-net is not allowed in NODE_OPTIONS, exit 9); this server is not sandboxed, so there is nothing to pass on. An oam that was itself started with --permission re-adds the flags inside spawn(), which the launcher cannot undo, so do not start it that way.
oam key support
oam 0.21.1 supports ed25519 and ECDSA key files, servers whose only host key is ed25519, and curve25519 key exchange in ssh2. Measured through the launcher on oam 0.21.1 (oam run bin/ssh-mcp.mjs): MCP initialize, tools/list (the same 21 tools as on Node) and tools/call on ssh_key_list, ssh_config_lookup and ssh_agent_ensure; and against a local ssh2 server, identical to Node 22, an ed25519 key file, an ECDSA key file, an ed25519-only host key and curve25519-sha256 all complete an exec, with ssh2's probes reporting eddsaSupported = true and curve25519Supported = true. An ed25519 key held in the agent was not measured end to end; what used to block it was ssh2 failing to parse an ed25519 public key on oam, and the underlying crypto.verify gap is closed.
The four node:crypto gaps that used to break these on oam 0.20.2 -- a crypto.verify that threw on an Ed25519 private key (which is how ssh2 decides it supports ed25519), a missing crypto.diffieHellman, x25519 key pairs that could not be generated, and SEC1 (EC PRIVATE KEY) signing that threw -- are listed in the launcher's comments and were tracked in YawLabs/oam#489, now closed.
To pin Node regardless of what the default prefers:
On Windows, add the same env block to the cmd /c form from Quick start.
Requirements
- Node.js 18+
- SSH client installed (for diagnostics and environment management)
- Optional: oam 0.21.1+ (preferred by
auto when a usable one is installed) — see Runtime selection
License
MIT
[Follow @YawLabs on X]