Provides SSH command execution tools for remote Unix/Linux server access
This SSH MCP server has 4 tools with basic schemas and descriptions, but significant quality gaps prevent a higher score. Tool naming is clear (ssh_execute, ssh_execute_script, ssh_upload_file, ssh_download_file), but descriptions lack LLM-optimized detail and parameter documentation is incomplete. The server provides input schemas with type definitions but missing critical details like parameter constraints, validation rules, and output schema documentation. Error handling is minimal, responses include generic error messages without recovery guidance. Security concerns exist around credential management (environment variables are better than hardcoding, but no mention of secure injection patterns). The code shows working SSH implementation but the tool definitions do not meet production quality baselines.
Download a file from the remote server via SFTP.
Execute a command on the remote server via SSH. Args: command: Shell command to execute timeout: Command timeout in seconds (default 30) Examples: - "ls -la" - "df -h" - "cat /proc/cpuinfo" - "python3 --version" Returns: Command output and status
Execute multiple commands as a script on the remote server. Commands are executed in sequence in a single SSH session. Args: commands: Newline-separated commands to execute timeout: Total timeout in seconds (default 60) Example: commands="cd /home/user\nls -la\npwd" Returns: Results of all commands
Upload a file to the remote server via SFTP.
Incomplete tool descriptions lack LLM decision guidance. ssh_upload_file and ssh_download_file have only 1-sentence descriptions (35 chars). 'Upload a file to the remote server via SFTP.' does not explain when to use this vs ssh_execute, what prerequisites exist, or what happens on error.
Output schemas are not documented. Tools return structured dicts (e.g., {success, stdout, stderr, exit_code} from ssh_execute) but descriptions do not specify what fields are returned or their types. LLMs cannot plan downstream actions without knowing response structure.
Error responses lack recovery guidance. When execute_command() catches an exception, it returns {success: False, error: str(e), ...} with a raw exception message. Per pattern:recovery-guide, errors must guide the LLM: 'Connection failed: [Errno 111] Connection refused' is not actionable, should be 'SSH connection failed (port 22, host may be down). Verify SSH_HOST and SSH_KEY_PATH environment variables.'
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 49 | <=2025-11-25 | v2 |
Parameter descriptions are minimal or missing constraints. 'timeout' parameters have descriptions like 'Command timeout in seconds (default 30)' but lack range info (e.g., '1-300 seconds'). 'command' parameter description in ssh_execute does not warn about shell injection risks or shell interpretation. LLMs cannot validate inputs without explicit constraints.
No tool annotations for safety. Tools like ssh_execute_script (execute arbitrary commands) and ssh_upload_file (write to remote filesystem) are destructive operations but have no destructiveHint or idempotentHint annotations. LLMs cannot determine which operations are safe to retry or require confirmation.
No confirmation or dry-run pattern for destructive operations. ssh_execute can run 'rm -rf /' and ssh_upload_file can overwrite critical remote files without any confirmation flow or dry-run capability. Per pattern:confirmation-request, irreversible operations should support a confirmation step to prevent catastrophic agent errors.
Missing multi-tool composition helpers. There is no ssh_list_files, ssh_read_file, or ssh_get_directory_contents tool. LLMs cannot explore the remote filesystem to understand context before running commands, forcing blind execution or multiple sequential calls.
Credentials exposed via environment variables without mention of secure injection. While using env vars is better than hardcoding, the code has no mention of secret vault integration, key rotation, or secure handling. The config.json fallback stores credentials in plaintext on disk (config.json). Per pattern:secret-injection, credentials must use secure injection and never appear in tool parameters (this server doesn't expose them as params, which is correct, but security posture is weak).