Kika's local-first MCP server for Obsidian vaults — direct filesystem access, works without the Obsidian app. Works with any MCP client (Claude Code, Claude Desktop, Cursor, Cline, Codex, Grok). Includes schema-validated Obsidian Bases (.base) support.
The server presents 17 tools with clear verb-based naming (configure_, get_, create_, update_, delete_, list_, search_). Most tools have basic descriptions and visible input schemas in server.py. However, there are significant quality gaps: (1) Many descriptions are generic and under 100 chars, lacking context on WHEN to use the tool or WHY vs. similar alternatives. (2) Several parameters lack type constraints (e.g., 'metadata' in update_note is just 'object' with no schema of allowed fields). (3) No output schemas are documented, LLMs cannot predict what fields will be returned. (4) Error handling is minimal, most functions return bare error strings rather than actionable guidance. (5) Security concerns: destructive operations (delete_note, delete_base) have no confirmation step or dry-run support. (6) The tool composition is reasonable (single responsibility per tool), but parameter descriptions often omit format constraints, ranges, or enums. This is a solid C-grade server, functional but not production-grade.
Configure the Obsidian vault path.
Create a new Obsidian Base (.base file).
Create a new folder in the vault.
Create a new note.
Delete an Obsidian Base (.base file) permanently.
Delete a note permanently.
Get all unique tags used in the vault.
No output schemas documented for any tool. LLMs cannot predict return field names, types, or structure, forcing them to reason about the response at runtime. This violates the pattern:tool requirement and increases hallucination risk.
No pagination support (limit, offset, cursor) on list_* and search_* tools. list_notes and list_bases can return unbounded results, risking context window exhaustion. Per the rubric baseline, list tools should cap at 20-50 items by default and offer pagination.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 59 | <=2025-11-25 | v2 |
Find all notes that link to the specified note.
Get an Obsidian Base (.base file) by path.
Get the complete folder structure of the vault.
Get a note by its path.
Get all wikilinks from a note.
List all Obsidian Base (.base) files in the vault or a specific folder.
List all notes in the vault or a specific folder.
Search notes by content, title, or tags.
Update an existing Obsidian Base (.base file) with partial changes.
Update an existing note.
Destructive operations (delete_note, delete_base) lack confirmation or dry-run support. Per pattern:confirmation-request, irreversible operations should offer a confirm step to prevent catastrophic agent mistakes.
Complex parameters (metadata, views, filters, formulas, properties, summaries) lack detailed schema definitions and descriptions. LLMs cannot infer allowed keys, types, or nesting. Per pattern:constrained-input, free-form objects should be documented with examples or enums.
Tool descriptions are generic and often under 60 characters. Missing context on WHEN to use vs. similar tools (e.g., get_backlinks vs. get_note_links distinction unclear). Per the rubric, descriptions should be 50-200 chars and include dependency hints.
Error handling is minimal, most tools return bare error strings (e.g., 'Note not found: {path}') with no recovery guidance. Per pattern:recovery-guide, errors should guide the LLM to next steps (e.g., 'Try search_notes() with partial name').
No tool annotations (readOnlyHint, destructiveHint, idempotentHint) present in the MCP tool definitions. These help clients flag dangerous operations and are part of the current spec (2026-07-28).
Parameter naming inconsistency: some params accept paths (path, folder_path, folder) without consistent naming convention. Per the rubric, parameter names should be suffixed with type hints (path vs. folder_name vs. folder_path).