Loading ToolBench
AST-based code analysis with DAG construction for semantic codebase search and documentation generation
This MCP server has critical definition quality gaps. Only 1 tool is defined (batch_search_codebase), and while it has a description and input schema, the schema has significant issues: parameters lack descriptions (only 'query' is documented), the 'k' parameter type is defined but not described, and there is NO documented output schema. The tool description itself is generic (41 characters) and does not explain WHEN to use this tool, WHAT it returns, or HOW to interpret results. No error handling guidance is provided. The tool operates on a semantic search system but provides no context on what 'semantic search' means, what embedding model is used, or what the returned results look like. Critical for LLM planning: the response schema is completely undocumented, the agent cannot know what fields to expect, making composition with downstream tools impossible.
Run parallel semantic searches across the codebase
No output schema documented. LLM cannot determine what fields the tool returns, breaking downstream composition and forcing agents to guess or extract from unstructured text.
Tool description is 41 characters: 'Run parallel semantic searches across the codebase', too vague. Does not answer WHEN to use this tool (vs. what?), WHAT it returns, or explain prerequisites (does it require indexed codebase?). LLMs cannot select this confidently.
Parameter 'k' (number of results) is typed as integer but has NO description. LLM does not know what range is valid, whether there is a max, or what happens if k=0. Unbounded numeric params invite absurd values.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | F | 46 | 2026-07-28+ | v2 |
| 2026-03-09 | F | 28 | - | v1 |
Parameter 'query' description 'Search query string' is generic and does not explain query syntax (free text, code patterns, regex?), length limits, or example patterns. LLMs have no guidance on crafting effective queries.
No error handling guidance. What happens if: codebase is not indexed? Query is invalid? k exceeds max results? Tool should describe recovery paths, but none are provided.
Tool accepts minItems=1, maxItems=5 array but does not explain what 'parallel' means, does order matter? Are results deduplicated across searches? Do LLMs need to await all 5 results or can they process streaming responses?