Loading ToolBench
AI-powered code analysis with MCP server — gives AI assistants deep understanding of your codebase with persistent memory across conversations
CodeAtlas provides 5 tools with basic descriptions and partial schemas. All tools are READ_ONLY and follow verb_noun naming (list_, get_, search_). However, critical gaps exist: parameter descriptions are minimal or absent (e.g., get_project_structure's 'type' and 'limit' lack descriptions), output schemas are completely undocumented, and error handling guidance is missing. The server lacks pagination support despite tools returning potentially large result sets. Descriptions are present but generic (avg ~80 chars), falling short of the 194-char production baseline. No tool annotations (readOnlyHint, idempotentHint) despite all being read-only operations.
Get import/call/containment/implements relationships between entities. Shows how modules, classes, and functions are connected.
Get AI-generated code insights including refactoring suggestions, security issues, and maintainability analysis.
Get all modules, classes, functions, and variables in the analyzed project. Returns entity type, name, file path, and line number.
List all projects that have been analyzed by CodeAtlas. Returns project names, paths, and last analysis time.
Search for functions, classes, modules, or variables by name. Supports fuzzy matching.
Output schemas completely undocumented. LLMs cannot infer what fields to expect from list_projects, get_project_structure, get_dependencies, get_insights, or search_entities. This forces agents to guess at response structure and breaks downstream tool chaining.
Parameter descriptions missing or incomplete. get_project_structure's 'type' and 'limit' parameters lack descriptions; get_dependencies' 'relationship' enum lacks explanation of what each type means (import vs call vs contains vs implements). LLMs cannot determine valid values or when to use each parameter.
No pagination support. Tools like get_project_structure and get_dependencies accept 'limit' but no offset/cursor/page parameters. Large codebases will return truncated results with no way to fetch remaining data. Violates paginated-result pattern.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | 2026-07-28+ | v2 |
No error handling guidance. Tools provide no recovery hints (e.g., 'Project not found. Call list_projects() to see available projects'). Agents cannot self-correct on failures.
No tool annotations despite all being read-only. Missing readOnlyHint and idempotentHint annotations that would signal to agents these are safe to call repeatedly without side effects.