Vault Graph MCP shows solid foundational work with clear naming conventions and documented schemas. All 8 tools follow verb_noun patterns (graph_*) and have explicit JSON Schema definitions with required parameters marked. However, parameter descriptions are inconsistent, some parameters lack detail on format/constraints, and output schemas are defined in TypeScript interfaces but not documented in the tool registration visible in the source. No tool-level error handling guidance or recovery hints are present. The server targets a well-defined domain (Obsidian vault graph navigation) and implements coherent composition, but lacks the polish needed for >70: descriptions could be LLM-optimized further, parameter constraints (e.g. limits on `limit`, format of `node` IDs) are not enforced or clearly explained, and no acknowledgment of edge cases or failure modes.

• Add explicit bounds to all numeric parameters in descriptions: 'limit (1 - 100, default 20)', 'force boolean (true/false)'. Update graph_get_neighbors_ranked, graph_get_related, graph_get_hubs, graph_search to state max results.
• Document output schemas inline with tool descriptions. E.g., for graph_get_neighbors_ranked, add: 'Returns array of ScoredNode objects with fields: id (string), score (number), explanation (string), commonNeighbors (number), degree (number), tagOverlapScore (number).'
• Add error recovery hints to descriptions. E.g., graph_get_node: 'If node is not found, try graph_search() to discover valid node IDs.' graph_build_index: 'If rebuild times out, try with smaller vaults or check vault format.'
• Specify node ID format precisely: 'Node ID is the file path relative to vault root without leading slash, including .md extension (e.g., "DevOps/GitOps.md", "README.md"). Case-sensitive.'
• Add composition hints to graph_search: 'Use this tool first to discover valid node IDs by title or tags, then pass node IDs to graph_get_neighbors, graph_get_node, or other navigation tools.'
• Consider adding optional 'node_title' parameter to graph_get_node, graph_get_neighbors, graph_get_neighbors_ranked to accept human-readable note titles as fallback (internally map to node ID via search if needed).
• Document what 'highlighted matched fields' means in graph_search output: does it return an array of matched terms, position offsets, or plain text with markup?
• Add guidance to graph_get_isolated_nodes on pagination: if limit is not provided or vault is large, implement cursor-based pagination or warn 'Large vaults may return truncated results; use limit parameter.'

Inferred effective spec: <=2025-11-25.

• Relies on Logging (deprecated) - log to stderr or use OpenTelemetry