Loading ToolBench
Developer intelligence MCP server for Claude Code — indexes your codebase, conventions, and review history into a vector database for RAG-powered code analysis
Tiburcio demonstrates solid tool naming, generally clear descriptions, and well-structured parameter schemas. All 12 tools follow verb_noun conventions (search*, get*, validate*). Descriptions are LLM-optimized (average ~140 chars, within the 10-1024 target). All input parameters have types and descriptions. However, output schemas are not explicitly documented in the provided source code, only input schemas are visible. Tool annotations (readOnlyHint) are declared, which is a strength. Error handling descriptions are minimal; most tools lack recovery guidance (e.g., 'if this fails, try X'). Composition is strong: each tool has a single responsibility, names are distinct, and parameters are well-constrained with enums. Security baseline is acceptable for a read-only tool set (all tools are marked READ_ONLY risk). One notable gap: getImpactAnalysis requires NEO4J_URI configuration but doesn't document failure modes if it's absent.
Search system architecture documents, flow diagrams, and component descriptions. Use when asked how components connect, data flows, or system design.
Catch up after time away. Returns a summary of changes over a specified period (e.g., 'what changed this week?').
Get comprehensive context for a file: conventions, recent review findings, and dependency information. Call this when starting work on an unfamiliar or historically problematic file, or before any refactor.
Blast radius analysis before refactoring. Shows which components and files depend on a given file or function. Requires NEO4J_URI to be configured.
Morning briefing of what changed overnight. Returns a summary of recent merges, review findings, and code changes from the nightly pipeline.
Output schemas not documented in source code, LLMs cannot predict return structure or plan downstream chaining.
Error handling lacks recovery guidance. Tools do not guide LLMs on what to do if a call fails (e.g., 'Try a broader query', 'Check if NEO4J_URI is configured').
getImpactAnalysis documents external dependency (NEO4J_URI) but provides no description of failure mode or how to handle it if missing.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | B | 70 | 2026-07-28+ | v2 |
| 2026-03-09 | C | 61 | - | v1 |
Retrieve a specific code pattern or boilerplate template by name. Use when you need ready-to-use code scaffolding for common patterns. If you don't know the pattern name, call without a name to list available patterns.
Get AI-generated test suggestions for recently changed code. Use when asked to write tests for recent merges.
Search real production code from the indexed codebase using hybrid search (semantic + keyword matching). Returns enriched results with symbolName, classContext, annotations, and exact line ranges. Use this to find existing implementations, patterns, and examples. For conventions and best practices, use searchStandards instead.
Search nightly code review notes from recent merges. Use when asked about recent changes, what was reviewed, or convention violations found.
Search database schemas, table definitions, and column descriptions. Use when asked about database tables, columns, or relationships.
Search your team's coding standards, conventions, and best practices. Use when you need to know HOW the team does something (transactions, error handling, batch jobs, Vue patterns, etc.). For actual code implementations, use searchCode instead. For code templates, use getPattern instead.
Check a code snippet against team conventions before committing. Makes an LLM call (~3-5s). Check validated:true in the response before trusting pass:true — pass:true alone may mean the check was skipped.
getNightlySummary and getChangeSummary descriptions under-specify their return semantics. 'Morning briefing' and 'Catch up' are vague; unclear whether they return structured data or plain text, how fresh the data is, or what to do if no changes occurred.
No pagination or result-limiting documented for search tools. Large codebases may return thousands of results; no guidance on max results, cursor/offset parameters, or truncation strategy.