One MCP server, any AI assistant — 6-8x fewer tokens on code reviews. Builds code graphs for efficient AI-assisted reviews using tree-sitter for multi-language support.
Server has 9 well-named tools with complete input schemas and reasonable descriptions. Naming follows verb_noun convention (build_graph, review_changes, get_impact, find_paths, search_symbols, get_symbol_details, get_file_symbols, export_graph, get_stats). All tools have descriptions 50-200 chars, meeting LLM-optimization baseline. However, output schemas are not documented, LLMs cannot predict return structure. Error handling is absent; no recovery guidance or categorization. Parameters lack validation constraints (e.g., max_depth unbounded, format strings unspecified). No tool annotations (readOnlyHint/destructiveHint). build_graph is marked WRITE but lacks confirmation/dry-run pattern for destructive operations.
Build or rebuild the code graph for a repository. Indexes all functions, classes, methods, and call relationships. Supports Python, JavaScript, TypeScript, and Go.
Export the code graph (json, dot, or summary).
Find all call paths from one symbol to another.
List all symbols defined in a specific file.
Find all upstream callers and downstream callees of a specific symbol.
Get statistics about the indexed code graph.
Get full details for a symbol: location, callers, callees.
Output schemas not documented. LLMs cannot predict return structure (fields, types, pagination). Breaks downstream tool chaining and forces agents to guess field names.
No error handling or recovery guidance. Errors will be raw exceptions with no actionable next steps for LLM. Missing error categorization (retryable vs user-fixable vs fatal).
build_graph (WRITE operation) lacks confirmation/dry-run pattern. Agents can accidentally rebuild entire codebase graphs without safeguard. No idempotency guarantee stated.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | <=2025-11-25 | v2 |
Given a list of changed files, return the minimal set of symbols and files the reviewer needs to look at. This is the core token-saving feature — instead of reading the whole codebase, your AI reads only the impacted blast radius.
Search for symbols by name or wildcard pattern.
Numeric parameters (max_depth, max_paths, limit) lack min/max bounds in descriptions. LLMs may pass absurd values (max_depth=999999) causing timeouts or resource exhaustion.
No tool annotations (readOnlyHint, destructiveHint, idempotentHint). Clients cannot distinguish safe read-only tools from destructive ones. build_graph should be marked destructiveHint=true.