Loading ToolBench
MCP Server exposing prompt search, organization, and RAG capabilities for managing a semantic prompt library with Copilot session processing.
This MCP server exposes 9 tools with well-structured JSON schemas and generally clear descriptions. However, there are gaps in parameter documentation, missing output schema specifications, and a few tools lack sufficient guidance for error recovery. The naming is consistent (verb_noun) and most tools are single-purpose, which is good. Parameters have type constraints and enums where appropriate. However, descriptions for several parameters are sparse, and output schemas are not explicitly documented in the tool definitions visible in the source code. The server includes READ_ONLY and WRITE classification in metadata, which aids security understanding. Missing: explicit output field documentation, comprehensive error handling guidance, and idempotent/destructive annotations in tool definitions.
Create a new prompt directly in the library without needing a Copilot session file. Auto-categorizes and indexes the prompt for immediate search.
Find prompts similar to a given prompt file
Get statistics about the prompt library (total prompts, categories, RAG status)
Get the full content of a specific prompt file
Get the full prompt library index with all sessions and metadata
Manually trigger RAG indexing of all prompts
List all prompts in a specific category
Output schemas not documented. Tools return results but LLMs cannot see the structure of responses (fields, types, nesting). This forces LLMs to guess what data is available and risks failed downstream tool chaining.
get_library_stats and get_prompt_index have minimal descriptions (60 chars or less) and lack detail on what fields they return or when to call them. LLMs may not understand when to use them vs other discovery tools.
Error handling guidance is absent from all tool descriptions. E.g., search_prompts does not document what happens if RAG is not initialized, what to do if no results are found, or what constitutes invalid input.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-21 | D | 57 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 21 | - | v1 |
Process and organize a session file exported from an AI coding agent (Copilot CLI, Claude Code, etc.). Extracts prompts, categorizes them, and indexes for search. Use 'create_prompt' for individual prompts instead.
Semantic search for prompts using RAG. Finds prompts by meaning, not just keywords.
Tool annotations (readOnlyHint, destructiveHint, idempotentHint) are present in metadata (Risk: READ_ONLY/WRITE) but not reflected in the MCP Tool definitions visible in source code. These should be part of the formal schema to guide LLM reasoning about safety.
No pagination support visible for list-like tools (list_prompts_by_category, get_prompt_index). Large libraries risk returning thousands of records, exhausting context and causing token overflow.
Path validation exists in code (mcp_server.py: _validate_safe_path) but is not documented in tool descriptions. LLMs may not understand path requirements, constraints, or examples.