Loading ToolBench
AI-agent-first codebase blueprint & dependency mapper for Python projects. Exposes CodeKnowledgeMapper analysis tools over the Model Context Protocol (MCP) using the SSE transport so AI agents can query code graphs, perform semantic searches, trace dependencies/dependants, and view component summaries.
CodeKnowledgeMapper exposes 8 read-only analysis tools with clear verb-noun naming (search_blueprint, get_node, get_impact, etc.). All tools have descriptions (avg ~120 chars, within baseline 34-392 range). However, input schemas are partially visible but lack critical details: parameter types are declared (string, integer) but descriptions are minimal or missing for several params. Output schemas are not documented. Error handling is absent, no guidance on what to do if a node_id is invalid or search returns empty. The server is HTTP-based (good), but lacks tool annotations (readOnlyHint, destructiveHint) and per-request _meta logLevel support. Thread-safety via RLock is well-designed but not a scoring factor.
Show all nodes that a given node depends on (calls/imports).
Impact analysis: show all nodes that depend on (call/import) a given node. Use this before making changes to understand what will be affected.
Retrieve the full blueprint entry for a specific node ID, including signature, file location, docstring, calls, and called_by list.
Get a compact AI-optimized summary of the codebase for fast orientation.
Get the nested component tree structure of the project.
List all modules in the project with their exports and dependencies.
Keyword & Domain-based functional search on the blueprint graph. Matches keywords inside names, IDs, signatures, docstrings, and AI summaries, filtering out unrelated business domains or namespaces.
Output schemas not documented. LLMs cannot plan downstream calls or extract required fields (e.g., what fields does search_blueprint return? Is there pagination?). Violates pattern:tool requirement.
No error handling or recovery guidance. If node_id is invalid or search returns empty, LLM receives no actionable error message. Violates pattern:recovery-guide.
Parameter descriptions are sparse or missing. 'limit' has a description, but 'query' in search_blueprint lacks format/constraint guidance. 'depth' in get_impact has no range (1-10? 1-100?). Violates pattern:tool-description.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 50 | <=2025-11-25 | v2 |
Manually trigger a fresh background scan and reload the store.
No pagination support. search_blueprint and list_modules return results but lack offset/limit or next_cursor. Large codebases could return hundreds of nodes, blowing context window. Violates pattern:paginated-result.
Tool annotations missing. No readOnlyHint on read tools or destructiveHint on trigger_scan. Agents cannot distinguish safe from risky operations. Violates current MCP spec (2026-07-28) tool annotation requirement.