Highly-performant MCP server for Obsidian vaults. Direct file access (no Obsidian app required), pre-computed indexes, in-memory state, sub-millisecond hybrid retrieval. Opinionated about a typed, atomic-claim, wikilink-graph schema.
Four tools with clear naming (verb_noun pattern) and detailed descriptions (150-300 chars each). Input schemas are well-structured with types and descriptions for most parameters. However, output schemas are not formally documented in the code, only described in docstrings. Parameter descriptions are thorough but lack explicit constraints (enums, ranges). Error handling guidance is minimal. wiki_reindex lacks input validation warnings.
Run 3-layer hybrid retrieval over the vault and return structured context. PRIMARY tool. Returns the most relevant atomic claims (with their source citations), graph-adjacent connecting concepts, and a short list of pages worth reading in full via wiki_read. The LLM does the synthesis from this context.
Fetch the full content of a vault page. Use AFTER wiki_query to drill into specific pages from the suggested_pages_to_read list. Accepts either a relative path (e.g. "Frameworks/Strategic Volume.md") or a page title (e.g. "Strategic Volume" — resolved via the graph index).
Manually trigger full reindex of the vault.
Return an overview of the vault's current state. Use to orient yourself before querying ("what's in this vault?") or to verify retrieval freshness (last_load_at timestamp).
Output schemas not formally documented. Tool descriptions mention return structure (e.g., 'dict with: query, keywords_used, filters_applied...') but no JSON Schema definition is visible in code. LLMs cannot reliably parse unstructured output.
wiki_query parameters lack explicit constraints. 'keywords' is described as 'Pass 3-5 variants' but no min/max array length enforced. 'max_claims' and 'max_pages' have no documented bounds (e.g., 1-100). LLMs may pass invalid values.
wiki_reindex has empty input schema (no parameters) but lacks documentation of side effects or confirmation pattern. Destructive operations should warn about irreversibility and offer dry-run or confirmation.
Error handling is implicit. No tool description explains what errors can occur, how to recover, or whether failures are retryable. E.g., wiki_read does not document 'page not found' behavior.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 57 | <=2025-11-25 | v2 |
wiki_read accepts 'page title' or 'relative path' but does not explain resolution order or failure modes. If both exist, which is preferred? What if neither matches?