Loading ToolBench
MCP server exposing 17 code intelligence tools for Claude Code, Cursor, Windsurf, and Zed. Builds a knowledge graph of your codebase using tree-sitter parsing and SQLite, enabling symbol lookup, dependency analysis, code validation, and AI hallucination detection.
Kodegraf exposes 17 code-intelligence tools with consistent naming (kodegraf_* prefix) and clear descriptions (avg 120 chars). All tools have descriptions and most have input schemas. However, critical gaps exist: (1) output schemas are undocumented, no tool declares what fields it returns, forcing LLMs to guess structure; (2) parameter descriptions are minimal or missing for many tools (e.g., kodegraf_impact 'files' param lacks format guidance); (3) no error handling guidance, tools don't explain recovery paths or categorize failures; (4) no pagination support despite tools returning potentially large result sets (kodegraf_search, kodegraf_flows, kodegraf_query); (5) tool composition is weak, no clear chaining guidance (e.g., after kodegraf_find, what's the next step?). Naming is strong (verb_noun pattern, action-oriented), but schema and output documentation are the primary blockers to production readiness.
Show which critical flows are affected by recent changes.
Build or rebuild the knowledge graph. Parses all source files and constructs the dependency graph.
Detect changed files since a git ref. Shows added, modified, and deleted files.
Validate a file against the knowledge graph. Catches wrong imports, non-existent functions, invalid enum values. Run this after writing code to verify correctness.
Find potentially dead code (functions/classes with no callers).
Show what depends on a file (what imports it). Use this to understand the impact of changes.
Output schemas completely undocumented. No tool declares return type, field names, or structure. LLMs cannot plan downstream calls or extract data reliably.
No pagination support. Tools returning lists (kodegraf_search, kodegraf_flows, kodegraf_query) lack limit/offset/cursor params and total_count in response. Large result sets will blow context windows.
Parameter descriptions are minimal or missing. 'files' in kodegraf_impact lacks format guidance (array of what? paths? globs?). 'pattern' in kodegraf_query is vague, does it accept regex or only predefined strings?
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2026-07-28+ | v2 |
Show what a file depends on (its imports). Use this to understand a module's dependencies.
Get the valid values of an enum. Use this before using enum values in code to avoid hallucinating invalid values.
List all exports from a module/file. Use this before importing from a file to know exactly what it exports.
Find functions, types, enums, classes, or any symbol in the codebase by name. Use this before writing code that references internal APIs.
List critical data/control flows in the codebase. Useful for understanding architecture.
Compute blast radius of changed files. Shows transitively affected files and dependencies.
Show quality insights and failure patterns from collected signals. Displays self-correction rate, top failure types, and recurring patterns.
Find large functions, classes, or files. Useful for identifying refactoring candidates.
Ultra-compact context (~100 tokens). Entry point for all tasks. Provides graph stats and suggests next tools based on task.
Query code relationships with predefined patterns: callers_of, callees_of, imports_of, importers_of, tests_for, file_summary.
Search the codebase knowledge graph for anything related to a topic. Returns matching functions, types, enums, modules.
No error handling guidance. Tools do not explain what to do on failure (e.g., if kodegraf_find returns no results, should the LLM try a broader search or a different tool?). No recovery paths documented.
Weak tool composition. No guidance on chaining (e.g., after kodegraf_find returns a symbol, what's the next step? kodegraf_query? kodegraf_deps?). LLMs must infer intent without explicit hints.