MCP server for safe remote SSH operations with plan approval, audit log, and credential management
This SSH MCP server defines 19 tools with reasonable naming and most tools have descriptions, but significant quality gaps prevent a higher score. Naming follows verb_noun pattern well (ssh_save_credential, ssh_execute_command, etc.). However, parameter schemas are partially visible in the spec but lack full JSON Schema type definitions in the source excerpt provided. Descriptions are present but many are generic (e.g., 'Execute a command on remote host' for ssh_execute_command is only 37 chars, below the 50-char baseline for A-grade tools). Critical issue: no visible output schema documentation for ANY tool, LLMs cannot reason about what these tools return. Error handling is minimal; tools do not guide recovery (e.g., 'Connection not found. Available connections: [...]'). Security model is strong (credentials not exposed as params, plan approval gates destructive ops), but this is offset by weak definition quality. Composition is good, tools are single-responsibility and chainable (e.g., ssh_create_plan → ssh_approve_plan → ssh_execute_plan). Overall, this is a functional but documentation-poor implementation.
Approve an execution plan with optional note.
Establish SSH connection using saved credential or direct parameters.
Create an execution plan for a command or file operation with approval tracking.
Delete a saved SSH credential.
Close SSH connection by name.
Download a file from remote host via SFTP.
Execute a command on remote host. High-risk commands require plan approval.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., what fields does ssh_list_connections return? Is there a connection_id?). This is a critical blocker for agent chaining.
Descriptions are too generic and often too short. Examples: 'List all active SSH connections' (30 chars, below 50-char baseline), 'List contents of a remote directory' (36 chars). These lack context on WHEN to use the tool, what it returns, or what to do next.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 53 | <=2025-11-25 | v2 |
Execute an approved execution plan and return results.
Retrieve audit log entries filtered by plan ID and/or status.
Retrieve details of a specific execution plan.
List all active SSH connections.
List contents of a remote directory.
List all execution plans with optional filtering by status.
List all saved SSH credentials.
Read a file from remote host.
Reject an execution plan with optional note.
Save SSH credentials for named connection. Supports private key path or key passphrase, not password.
Upload a local file to remote host via SFTP. Requires plan approval.
Write content to a file on remote host. Requires plan approval.
Error handling is absent. Tools do not return actionable recovery guidance. Example: if ssh_connect fails because the connection_name does not exist, does the tool suggest available credentials? If ssh_execute_command is rejected due to high-risk detection, does the error guide the user to create a plan instead?
Parameter descriptions lack constraint information. Examples: ssh_execute_command has a 'timeout' param with no min/max documented (should state e.g., '1-3600 seconds'). ssh_write_file has 'permissions' param described as 'File permissions (octal)' but no format example (should be '0644, 0755, etc.'). ssh_create_plan has 'kind' param with enum constraint in description ('command, file_edit, upload, or key_auth') but this should be formalized in JSON Schema enum.
No pagination documented for list operations. ssh_list_saved_credentials, ssh_list_connections, ssh_list_plans, and ssh_list_directory do not specify limit, offset, or total count. If there are 1000+ credentials, LLMs cannot paginate results, risking context overflow.
ssh_save_credential accepts 'jump_host' as an object but does not document its schema. What fields must it have? Is it hostname, username, port, or something else? LLMs cannot construct valid jump_host payloads without this.