Sandboxed filesystem MCP server speaking Streamable HTTP, for ChatGPT connectors
local-mcp demonstrates solid fundamentals: all 9 tools have clear verb-noun naming (read_file, write_file, execute, etc.) and descriptions ranging 50 - 150 chars. Input schemas are fully visible with typed parameters and descriptions. However, output schemas are undocumented, LLMs cannot predict what fields to expect from responses. Error handling is minimal; tools lack recovery guidance. Parameter descriptions are functional but sparse (e.g., 'Path relative to the sandbox root' lacks format/constraint detail). No tool annotations (readOnlyHint, destructiveHint) despite clear risk levels. Composition is sound: each tool has one responsibility, and naming makes distinctions clear (read_file vs write_file vs edit_file).
Replace one exact occurrence of old_text with new_text. Fails if old_text is missing or appears more than once.
Run a shell command and wait for it. If it is still running after the timeout, returns a job_id to poll instead.
List directory entries with sizes.
Poll a running job for its current status and output.
Read a text file, returned with line numbers.
Search file contents by regular expression, honouring .gitignore.
Output schemas are completely undocumented. LLMs cannot predict response structure (fields, types, pagination). This forces agents to guess what data is available and risks context loss between tool calls.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) despite clear risk levels. execute, start_command, stop_job are IRREVERSIBLE; write_file and edit_file are WRITE. Annotations help agents reason about safety and retry logic.
Error handling lacks recovery guidance. Tools return errors but do not tell LLMs what to do next (retry, ask user, abort). E.g., 'File not found' should suggest 'Call list_dir() to see available files.'
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 64 | <=2025-11-25 | v2 |
Start a shell command in the background and return its job_id.
Stop a running job.
Create a file or replace its entire contents.
Parameter descriptions lack constraint detail. 'Path relative to the sandbox root' does not specify format, length limits, or allowed characters. 'Rust regular expression' in search() assumes LLM knows Rust regex syntax.
execute() and start_command() accept arbitrary shell commands with no validation or sandboxing hints. Description does not clarify what commands are allowed, what environment variables are available, or what the timeout behavior is.