An MCP server that provides secure SSH command execution, file reading, log tailing, and job management with comprehensive authorization, audit logging, and credential masking.
SshWarden defines 7 tools with complete input schemas and descriptions. All tools have verb-noun naming (list_changes, start_job, poll_job, kill_job, read_file, tail_log, run). Descriptions are substantive (100-300 chars), explaining WHAT the tool does and WHEN to use it. Parameters are typed and described. However, output schemas are not documented, LLMs cannot predict return structure. Error handling is mentioned in descriptions but not formalized with recovery guidance. No tool annotations (readOnlyHint/destructiveHint) despite clear risk levels. Risk classifications exist (READ_ONLY, WRITE, REVERSIBLE) but are not exposed in the MCP schema.
Stop a running job by sending a signal to its process group on the target.
List files under this deployment's watched paths that were created, modified or deleted on a host recently. Changes are noticed by a periodic sweep, so the resolution is the sweep interval: a change made and undone between two sweeps is not seen, and neither is one that leaves size, modification time and inode alone.
Read a job's status and whatever it has printed since a given line. Status is one of running, finished, gone (not running and left no exit status - signalled, or the machine restarted) or vanished (its directory is no longer on the target).
Read a file on a host. Which files are readable is decided by this deployment's configuration and by what the unix account it maps you to can open. Every call is recorded. Credential-shaped values are masked before the content is returned.
Run a shell command on a host over SSH and wait for it to finish. Every call is independent: nothing is remembered between calls, so the working directory and any environment variables must be given each time. Every call is recorded.
Output schemas not documented. LLMs cannot predict return structure for any tool. Descriptions mention 'returns a job id' or 'returns status' but do not specify field names, types, or structure.
Tool annotations missing. Risk levels (READ_ONLY, WRITE, REVERSIBLE) are tracked internally but not exposed via MCP readOnlyHint/destructiveHint/idempotentHint. LLMs cannot reason about side effects.
Error handling lacks recovery guidance. Descriptions mention 'credential-shaped values are masked' and 'paths containing .. are refused' but do not tell LLMs what to do when these constraints are violated.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | <=2025-11-25 | v2 |
Start a command on a host and return immediately with a job id. The command keeps running after this call returns and after this server restarts. Use poll_job to read its output and kill_job to stop it. Its output is written to a private file on the target, and is masked for credential-shaped values only on the way back through poll_job.
Read the end of a log on a host - either a service unit's journal or a log file. An argument beginning with '/' is read as a file path; anything else is read as a service unit name. Every call is recorded, and credential-shaped values are masked.
Parameter constraints not formalized. 'maxBytes' and 'lines' are integers with no min/max bounds documented. 'signal' in kill_job accepts free-form strings (e.g., TERM, KILL) but no enum constraint.
No pagination support. list_changes returns 'files under watched paths' with no limit, offset, or total count. Large deployments could return thousands of items, exhausting context.