MCP SSH Server - Connect to remote hosts via SSH through MCP protocol
mcp-ssh has registered tools with basic names and descriptions, but significant quality gaps prevent higher scoring. Tool names follow verb_noun conventions (ssh_connect, ssh_execute, ssh_close), which is good. However, input schemas visible in the code are INCOMPLETE, the provided source shows tool registration WITHOUT explicit InputSchema objects in the mcpsdk.AddTool() calls. The Makefile and go.mod are shown, but the actual handlers (mcp/handlers.go) containing schema definitions are not provided. Tool descriptions are present (ranging 50-100 chars) but lack depth required for LLM planning, they do not explain WHEN to use each tool, WHAT prerequisites exist, or HOW the tool chains with others (e.g., ssh_connect is required before ssh_execute, but this is not stated). Parameter descriptions are minimal or missing entirely in the visible code. No output schemas are documented. Error handling is likely present in handlers (not shown) but cannot be verified. Security concerns are HIGH: ssh_connect accepts password and private_key_path as parameters, which violates the secret-injection pattern, these should never be tool parameters. Overall per-tool average: 42/100.
Close an active SSH connection
Establish an SSH connection to a remote host
Execute a command on an active SSH connection. Environment variables and working directory persist between commands.
Execute a command asynchronously and get a job ID for polling status
Cancel a running job
List all jobs for a connection
Get the status of an asynchronous job
CRITICAL: Input schemas not visible in provided source code. Tool definitions show parameter names in the evaluation metadata, but mcpsdk.AddTool() calls in main.go do NOT include InputSchema objects.
CRITICAL: Credentials exposed as tool parameters. ssh_connect accepts 'password' and 'private_key_path' as input parameters. Credentials must NEVER appear as tool parameters, they are logged in traces and prompt history. Use server-side secret injection via environment variables or vault instead.
HIGH: Parameter descriptions incomplete or missing. The visible schema metadata shows parameters like 'max_lines', 'max_bytes', 'use_login_shell', 'enable_pty' with descriptions, but these lack format/range guidance. E.g., max_lines has no minimum/maximum constraints; enable_pty has no explanation of when PTY allocation is needed (interactive vs batch work).
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 40 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 43 | - | v1 |
List all active SSH connections
HIGH: Tool descriptions lack dependency guidance and prerequisites. ssh_execute says 'Execute a command on an active SSH connection' but does NOT state 'Requires ssh_connect to be called first' or explain how connection_id must match a prior connection. LLMs cannot infer multi-step dependencies.
MEDIUM: No output schemas documented. Tool descriptions do not explain what ssh_execute returns, what fields are in ssh_list results, or what ssh_job_status provides. LLMs need explicit return type documentation to plan downstream tool calls.
MEDIUM: Async job tools lack result limits and pagination. ssh_job_list accepts 'connection_id' only, no limit/offset parameters visible. If a connection has 1000 jobs, returning all could blow context window. No documentation of max result count.
MEDIUM: Missing error handling guidance. Handler implementations not shown, but tool descriptions do NOT indicate what to do on common failures (e.g., connection_id not found, command timeout, permission denied). LLMs need 'recovery guides' to know whether to retry, escalate, or ask the user.
MEDIUM: Tool naming ambiguity in async API. ssh_execute and ssh_execute_async differ only in '_async' suffix, which is implementation-detail leakage. Better names: ssh_execute_sync (explicit blocking) and ssh_execute_background (explicit async) or ssh_execute_fire_and_forget. Current names force LLMs to reason about when to use which variant.