Obsidian MCP Server

obsidian-mcp-server

A Go Model Context Protocol (MCP) server for Obsidian vaults. It gives AI agents and development tools direct filesystem access to your vault — no running Obsidian instance required.

Features

• Read, write, and list notes and directories via MCP tools
• Frontmatter — parse and update YAML frontmatter with format-preserving rewrites
• Tags — extract inline #tags, aggregate vault-wide tag counts, add/remove tags
• Backlinks — on-demand reverse link graph (wikilinks and markdown links)
• Mutations — heading-anchored patch, safe delete, and move with confirmation guards
• Full-text search — BM25 Okapi ranked search with match snippets
• Regex/glob search — RE2 regex or filepath glob search across paths and content
• Path security — 4-layer validation: lexical checks, ignore/extension filters, case-insensitive existence lookup, and symlink escape prevention
• Stdio transport — works with any MCP client (Claude Code, Claude Desktop, etc.)
• Streamable HTTP transport (optional) — TLS 1.3 + bearer token auth by default, loopback-only unless explicitly widened; see --transport http in Configuration and SECURITY.md § HTTP transport
• Zero Obsidian dependency — operates on the vault directory directly
• Token counting — responses include approximate token counts (cl100k_base)

MCP Tools

search_notes, search_regex, list_directory, and get_recent_changes results also include an obsidian://open deepLink field per note, built from --vault-name (see Configuration below).

Notes

`patch_note` semantics: position controls where content is inserted relative to the heading:

• before — inserted immediately before the heading line
• after — inserted after the heading's body (before the next same-level or higher heading)
• replace_body — replaces everything between the heading line and the next same-level heading

`search_notes` parameters:

Returns: { query, results: [{ path, score, matchCount, matches: [{line, snippet, term}], tokenCount, reason }], total }

`search_regex` parameters:

Returns: { pattern, scope, results: [{ path, matches: [{line, snippet}] }], total }

Batch tools (`read_multiple_notes`, `get_notes_info`): The paths parameter is a JSON array string — e.g. '["Notes/foo.md","Notes/bar.md"]'. summary:true returns headOf (first N runes from headChars, default 200) instead of full content, which is useful for large notes to stay within context limits. Both tools enforce --max-batch (default 10); requests with more paths are silently truncated and the response includes "truncated": true.

Optimistic concurrency (`if_match` / `etag`): read_note, read_multiple_notes, and get_notes_info return a SHA-256 etag of the note's content. Pass that value as if_match on write_note, patch_note, update_frontmatter, manage_tags, delete_note, or move_note to make the write conditional — if the note has changed since you read it, the call fails with a REVISION_CONFLICT error instead of silently overwriting someone else's edit. if_match is optional everywhere; omitting it writes unconditionally, as before. Two edge cases: passing if_match for a note that doesn't exist yet is always a conflict (it never creates the note), and move_note's dryRun:true preview does not enforce if_match.

Periodic notes (`get_periodic_note`, `get_recent_periodic_notes`): Configuration (folder and date format per granularity) is read from .obsidian/plugins/periodic-notes/data.json inside the vault. If that file is missing, built-in defaults are used: daily notes use YYYY-MM-DD in Daily Notes/, weekly notes use gggg-[W]ww in Weekly Notes/, and so on. offset=0 resolves to the current period, offset=-1 to the previous period (yesterday, last week, etc.), and offset=+1 to the next period. createIfMissing=true creates an empty note at the resolved path if it does not already exist.

`get_vault_stats`: Returns noteCount, totalBytes, totalLinks, totalTags, topTags (top 20 by count), oldestNote, newestNote, and vaultRoot. Setting includeTokenCounts:true runs token counting across every note — this is expensive for large vaults and is disabled by default.

`audit_notes` classes:

• orphans — notes that have no tags AND no incoming wikilinks or markdown links (completely isolated notes)
• dangling-links — notes containing links to vault paths that do not exist (broken references)
• untagged — notes with no frontmatter tags and no inline #tags
• duplicate-titles — multiple notes sharing the same filename stem, which causes wikilink ambiguity

Each class result is capped at limit entries (default 20). When results are truncated, the response includes "truncated": true.

MCP Prompts

Prompts are server-defined conversation starters that the host (Claude Code, Claude Desktop) exposes in its UI. Each prompt pulls live vault data and constructs a ready-to-use message for the LLM.

Prompts are invoked from the host's prompt picker (e.g. / in Claude Code). They never modify the vault.

MCP Resources

Resources are read-only vault data that the host can attach directly to a conversation context window — no explicit tool call required.

| Vault statistics | `obsid