SSH/SFTP MCP Server - Execute remote commands and transfer files via SSH for Claude AI
The server defines 18 tools with mostly complete input schemas and descriptions, but with notable gaps in output schema documentation, parameter constraints, and error handling guidance. Tool naming follows verb_noun conventions consistently (ssh_*, sftp_*). Descriptions are present but vary in quality and specificity. Critical security concern: passwordOverride and direct password parameters expose credentials in logs. Schema definitions use Zod with types and descriptions, which is good. However, output schemas are not documented, LLMs cannot predict response structure. Error handling is basic (success/error JSON) but lacks recovery guidance or actionable instructions. No tool has enums, ranges, or format constraints documented. The server lacks pagination, filtering beyond basic tag filtering, and result limits despite potential for large responses (e.g., sftp_ls).
Download a file from a remote server via SFTP.
List directory contents on a remote server via SFTP.
Create a directory on a remote server via SFTP.
Read a file from a remote server via SFTP.
Remove a file or directory on a remote server via SFTP.
Get file/directory statistics on a remote server via SFTP.
Upload a file to a remote server via SFTP.
Write a file to a remote server via SFTP.
Create an SSH connection session with provided credentials.
Credentials exposed as tool parameters: 'password', 'passwordOverride', 'privateKey', 'passphrase', and 'sudoPassword' all appear as parameters. These will be logged in agent traces, system logs, and prompt history, creating a security breach.
Output schemas are not documented. LLMs cannot predict the structure of responses (e.g., what fields ssh_test_connection returns). This prevents accurate downstream tool composition and forces agents to guess at response structure.
No pagination or result limits on potentially large responses. sftp_ls could return thousands of files in a directory; ssh_list_servers could return hundreds of servers. Without pagination and limits, responses can exhaust context windows.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 50 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 45 | - | v1 |
Connect to a server using its configuration ID. Credentials are loaded from config.
Close an SSH session.
Execute a command on a remote SSH server.
Get details of a specific configured server by ID.
List all configured SSH servers. Servers can be defined in servers.json or via environment variables.
List all active SSH sessions.
Execute a command with sudo privileges on a remote SSH server.
Test connections to all configured servers and report status.
Test SSH/SFTP connection to a configured server without creating a session.
Error responses are generic JSON structures (success: false, error: string). They do not guide LLM recovery, no indication whether to retry, ask the user, or treat as fatal. E.g., 'Server not found' doesn't suggest calling ssh_list_servers first.
No enum constraints or parameter value validation documented. 'encoding' in sftp_read/sftp_write accepts any string; 'timeout' in ssh_exec is unbounded; 'mode' in sftp_mkdir is expressed as octal but not constrained in the schema. LLMs can pass invalid values.
No idempotency guarantees documented for stateful operations. Agents retry on ambiguous failures; sftp_upload with overwrite=false, sftp_write, and ssh_exec may produce duplicate side effects if retried. No documented idempotent markers.
No confirmation or dry-run support for destructive operations. sftp_rm can recursively delete directories; ssh_exec can run arbitrary commands. Agents can accidentally invoke these with wrong parameters, no confirmation gate.
Parameter descriptions lack format/constraint details. 'privateKey' doesn't specify whether it expects a file path or key content; 'remotePath' in sftp_ls doesn't indicate whether it's required to exist or if a non-existent path returns an error vs empty list.