MCP server for filesystem operations over SSH using asyncssh
This server exhibits significant gaps in definition quality. While most tools have descriptions (90%+ present), descriptions are often generic and lack actionable context for LLM selection. Schemas are present for basic string parameters but lack sophisticated constraints (enums, ranges, formats). No error handling guidance is visible. Naming conventions are generally good (verb_noun pattern), but parameter descriptions are sparse or missing. The server conflates duplicate tool names (e.g., three 'read_text_file', three 'list_directory', two 'list_markdown_sessions') across different backends without clear disambiguation, creating confusion for LLMs selecting between them. No tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present. Output schemas are not documented. The 'sync_markdown_to_trilium' and 'sync_directory_to_trilium' tools accept secrets (trilium_token) as parameters, critical security violation. See src/mcp_filesystem_sshfs/server.py for tool registration.
Create a directory on the remote server (creates parent directories as needed).
Return a machine-readable description of available tools, their parameters and examples.
Get detailed metadata about a file or directory on the remote server.
Return the connector configuration for a given store_id.
List all directories that this server is allowed to access on the remote server.
List directory entries for `path` on the specified store.
List contents of a directory on the remote server.
Secrets exposed as tool parameters: 'trilium_token' is passed as a string parameter in sync_markdown_to_trilium and sync_directory_to_trilium. This violates pattern:secret-injection, secrets in tool parameters are logged, stored in traces, and exposed in agent audits.
Duplicate tool names across backends without clear disambiguation. Three tools named 'read_text_file', three named 'list_directory', two named 'list_markdown_sessions' are registered across single-store and multi-store backends. LLMs cannot distinguish between them and will make arbitrary selection choices, causing unpredictable behavior.
Parameter descriptions are minimal or missing. Most parameters have 1-2 word descriptions ('Path to the file on the remote server'). No documentation of valid ranges, formats, constraints, or dependencies. Param 'pattern' in search_files lacks description of wildcard syntax semantics.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
List Markdown files under a remote directory using SFTP.
List Markdown files (sessions) under `remote_dir` on the given store.
Return a list of configured stores.
Move or rename a file or directory on the remote server.
Read the full contents of a remote Markdown file and return it as UTF-8 text.
Read a Markdown file from a store and return its content.
Read a text file from `store_id` at `path` and return UTF-8 text.
Read complete contents of a text file from remote server.
Search for files matching a pattern in a directory tree on the remote server.
Sync Markdown chat transcripts into Trilium using idempotent mapping data.
Read a Markdown file and create a Trilium note for it (returns note info dict).
Create or overwrite a file on the remote server.
No output schema documentation. Tools return formatted strings (e.g., 'list_directory' returns '\n'.join(lines)) but nowhere is the structure of returned data documented. LLMs cannot plan downstream calls or parse results reliably.
No error handling or recovery guidance in tool descriptions. If a file is not found, permission denied, or SSH connection fails, the agent receives a raw exception with no actionable next steps (pattern:recovery-guide).
No input validation described. Parameters like 'path' and 'pattern' accept arbitrary strings with no validation rules, ranges, or format constraints documented. LLMs cannot infer what is valid and will pass malicious or nonsensical inputs.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). LLMs cannot distinguish which operations are safe to retry, which are irreversible, or which are read-only. This prevents proper error recovery planning.
Missing permission/scope declarations. Tools that modify files (write_file, create_directory, move_file) do not declare required scopes (e.g., 'write:filesystem'). Cannot audit which agent permissions are needed or apply least-privilege configurations.
Tool composition concerns: 'sync_markdown_to_trilium' and 'sync_directory_to_trilium' bundle file reading, Trilium API calls, and mapping logic into single tools. Should be split into read_markdown + send_to_trilium to allow agent composition flexibility.
No pagination or result limits documented. 'list_directory' and 'search_files' return unbounded results via string concatenation. Large directories or search results will blow context windows. No mention of limits, pagination, or offsets.