Loading ToolBench
MCP server for retrieving and applying coding rules for AI agents
The Rule MCP Server has 5 tools with basic schemas and descriptions, but exhibits significant quality gaps. Tool names are reasonably verb-based (getRules, validateCode, getProjectInfo, autoDetectProject, scanLocalProjects), but descriptions are terse and lack LLM-optimization guidance. Input schemas are present but lack descriptions for most parameters, critically, the 'language' parameter in getRules and validateCode has a description but project_id does not explain its format or where to obtain it. Output schemas are not documented at all, forcing LLMs to guess what fields to expect and how to chain calls. Error handling is minimal, no guidance on recovery paths, retryability, or what to do if a project_id is invalid. The server lacks resource context: a user saying 'analyze my code' cannot easily map to a project_id without additional discovery tools. Overall structure is functional but under-optimized for agentic reasoning.
Automatically detect project from path and get appropriate rules
Get information about a specific project
Get coding rules for a specific project
Scan local directory to detect multiple projects
Validate code against project rules
No output schemas documented. LLMs cannot plan downstream tool chains or extract expected response fields. For example, after validateCode returns violations, does it include line numbers, severity levels, and fix suggestions? Unknown.
Parameter descriptions are sparse or missing. 'project_id' (required in getRules, validateCode, getProjectInfo) has no description, LLMs don't know where to source it, its format, or how it maps to real projects. 'language' is described as '(optional)' but not its enum values or default behavior.
No error guidance or recovery paths. If validateCode fails (e.g. project_id not found, code syntax invalid), what should the LLM do? No error classification, no hints for retry vs user-fixable vs fatal.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 48 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 46 | - | v1 |
autoDetectProject and scanLocalProjects operate on filesystem paths. No validation guidance: do they accept absolute paths only, relative paths, or both? No security context for path traversal risk or sandboxing constraints.
validateCode tool description does not state whether it modifies state, only validates, or if it saves results. LLMs need to know if the call is safe to retry or has side effects.
Tool descriptions are under 20 characters on average (e.g. 'Get coding rules for a specific project' = ~39 chars, acceptable, but missing WHEN to use and prerequisite context). LLMs must know: call getRules before validateCode? Or are they independent?