Minimal MCP server exposing a sandboxed ./workspace folder over streamable HTTP. Transport-agnostic: any MCP client can use it (ChatGPT via the OpenAI Secure MCP Tunnel, Claude Code, Gemini CLI, Codex, Cursor, your own SDK code, ...). Includes path resolution guardrails and optional bearer token authentication.
workspace-mcp provides 5 tools with clear verb-noun naming (list_files, read_file, write_file, search_text, ping). All tools have descriptions and input schemas with types. However, descriptions are terse (10-60 chars), parameter descriptions lack detail about constraints and formats, output schemas are undocumented, and error handling is minimal. The ping tool is deprecated per MCP spec 2026-07-28. No tool annotations (readOnlyHint/destructiveHint) despite clear risk profiles. Schemas are present but sparse, no enums, ranges, or validation guidance. This is a functional but minimalist implementation.
List files and folders inside the workspace. `path` is relative to the workspace root.
Smoke-test tool: confirms the tunnel reaches this MCP server.
Read a UTF-8 text file from the workspace.
Search for a literal string in text files under the workspace.
Create or overwrite a UTF-8 text file in the workspace.
Descriptions are too brief (10-60 chars). LLMs need 50-200 char descriptions explaining WHAT, WHEN, and any prerequisites. E.g., read_file says 'Read a UTF-8 text file from the workspace' but omits max file size, encoding behavior on invalid UTF-8, or when to use vs list_files.
Parameter descriptions lack constraint details. 'path' is described as 'relative path to the workspace root' but omits: max length, allowed characters, whether '..' is blocked, or what happens if path escapes workspace. search_text's max_results lacks range (1-1000?) or rationale for default 50.
No output schemas documented. Callers cannot know what fields to expect. read_file returns a string, but is it truncated at MAX_READ_BYTES? search_text returns newline-delimited results, is the format stable? list_files returns 'dir' or 'file' labels, what if a symlink exists?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 55 | 2026-07-28+ | v2 |
No tool annotations despite clear risk profiles. write_file is destructive (overwrites files) but lacks destructiveHint. list_files and read_file are read-only but lack readOnlyHint. search_text is idempotent but lacks idempotentHint. Agents cannot reason about retry safety or side effects.
ping tool is deprecated in MCP spec 2026-07-28 (removed from core protocol). Agents should not rely on it for health checks. Replace with a resource or remove entirely.