Obsidian Semantic Search MCP

Read-only semantic retrieval for agents that need to find the right Obsidian note without write access.

한국어 · Quick Start · Why This Exists · How It Works

Your Obsidian vault is useful only if your agent can find the right note.

Keyword search misses context. Full write-capable Obsidian MCP servers expose more power than a retrieval agent needs. Obsidian plugins are great inside Obsidian, but they are not always the right boundary for Codex, Claude Desktop, Cursor, or any other MCP client.

This project is the narrow version:

local Obsidian vault -> read-only scanner -> local SQLite index -> MCP search/read tools

No note writes. No cloud embeddings by default. No Obsidian plugin runtime. No sync service.

Status: 0.3.1 early preview. The server is usable today, but ranking behavior and tool schemas may change before 1.0.

What You Get

Example result shape:

{
  "path": "02_Projects/RealtimeAPI/05_Interview_QA.md",
  "title": "Interview Q&A",
  "score": 0.7431,
  "matched_sections": [
    {
      "heading": "Level 4 > Redis Lua atomicity",
      "lines": [266, 305],
      "reason": "semantic=1, keyword=0.5565, metadata=0.6"
    }
  ]
}

Quick Start

Requirements:

• Node.js >= 24
• Ollama
• An Obsidian vault
• An MCP client such as Codex, Claude Desktop, Cursor, or another stdio MCP client

Install the embedding model:

ollama pull bge-m3
curl http://localhost:11434/api/tags

Print setup guidance:

npx -y --package @dalecb/obsidian-semantic-mcp obsidian-semantic-mcp-setup

Codex Setup

Add this to ~/.codex/config.toml:

[mcp_servers.obsidian_semantic]
command = "npx"
args = ["-y", "@dalecb/obsidian-semantic-mcp"]

[mcp_servers.obsidian_semantic.env]
OBSIDIAN_VAULT_ROOT = "/path/to/your/Obsidian Vault"
OBSIDIAN_SEMANTIC_MCP_HOME = "/Users/you/.obsidian-semantic-mcp"
OLLAMA_BASE_URL = "http://localhost:11434"
OBSIDIAN_EMBED_MODEL = "bge-m3"
OBSIDIAN_SEMANTIC_AUTO_INDEX = "true"

Restart Codex, then run:

obsidian_semantic.index_status
obsidian_semantic.search_notes { "query": "Redis Lua atomicity", "limit": 5 }

Claude Code Setup

Add the server at user scope so the personal vault path stays out of project .mcp.json files:

claude mcp add obsidian_semantic --scope user \
  --env OBSIDIAN_VAULT_ROOT="/path/to/your/Obsidian Vault" \
  --env OBSIDIAN_SEMANTIC_MCP_HOME="/Users/you/.obsidian-semantic-mcp" \
  --env OLLAMA_BASE_URL="http://localhost:11434" \
  --env OBSIDIAN_EMBED_MODEL="bge-m3" \
  --env OBSIDIAN_SEMANTIC_AUTO_INDEX="true" \
  -- npx -y @dalecb/obsidian-semantic-mcp

claude mcp get obsidian_semantic

Start a new Claude Code session and run /mcp to confirm the server is connected. On native Windows, use -- cmd /c npx -y @dalecb/obsidian-semantic-mcp after the environment flags.

See Anthropic's Claude Code MCP documentation for scope and command details.

Claude Desktop / JSON MCP Client Setup

In Claude Desktop, open Settings → Developer → Edit Config. This opens ~/Library/Application Support/Claude/claude_desktop_config.json on macOS or %APPDATA%\Claude\claude_desktop_config.json on Windows. Merge the following server into mcpServers:

{
  "mcpServers": {
    "obsidian_semantic": {
      "command": "npx",
      "args": ["-y", "@dalecb/obsidian-semantic-mcp"],
      "env": {
        "OBSIDIAN_VAULT_ROOT": "/path/to/your/Obsidian Vault",
        "OBSIDIAN_SEMANTIC_MCP_HOME": "/Users/you/.obsidian-semantic-mcp",
        "OLLAMA_BASE_URL": "http://localhost:11434",
        "OBSIDIAN_EMBED_MODEL": "bge-m3",
        "OBSIDIAN_SEMANTIC_AUTO_INDEX": "true"
      }
    }
  }
}

Fully quit and restart Claude Desktop. Cursor and other JSON-style stdio MCP clients can use the same mcpServers entry. On Windows, if npx is not found, use "command": "cmd" and prepend "/c", "npx" to args.

See the official local MCP server guide for Claude Desktop configuration and logs.

Why This Exists

This isn't aiming to be the most powerful Obsidian automation server. It aims to be the safest retrieval tool you can hand an agent.

Here's how it stacks up against the two tools it usually comes down to — a full-permission Obsidian MCP server (Local REST API based) and GBrain (a broader knowledge-compilation platform):

That trade is on purpose: give up writing, editing, and running inside Obsidian, and you get fewer moving parts and a smaller blast radius in return.

Use this if your agent should answer:

• "Which note explains this project decision?"
• "Find the file where I wrote about idempotency payload mismatch."
• "Show me the career notes related to this interview topic."
• "Search my vault, but do not mutate it."

Do not use this if you want an Obsidian UI plugin, automatic note generation, or write-capable vault automation.

Tools

index_status

Returns index metadata and safety settings.

index_vault

Builds or updates the external SQLite index.

{ "mode": "incremental" }

Specific files:

{
  "mode": "incremental",
  "paths": ["02_Projects/My Note.md"]
}

search_notes

Searches notes with hybrid semantic and keyword ranking.

{
  "query": "live coding notes",
  "limit": 8,
  "mode": "hybrid"
}

Modes:

• hybrid: semantic vector + SQLite FTS5 + metadata boosts
• semantic: vector-first search
• keyword: FTS5 keyword search without embedding the query

read_note

Reads a note or line range by vault-relative path.

{
  "path": "02_Projects/My Note.md",
  "start_line": 10,
  "end_line": 40
}

How It Works

index_vault
  -> scan Markdown files under OBSIDIAN_VAULT_ROOT
  -> block denied paths and symlink escapes
  -> split notes by Markdown headings
  -> create one summary chunk per file
  -> embed chunks with Ollama bge-m3
  -> store notes, chunks, FTS rows, and vectors in SQLite

search_notes
  -> incrementally index on first search or first search after a Markdown change
  -> embed the query with Ollama
  -> score vector similarity
  -> score SQLite FTS5 keyword matches
  -> apply title/path/heading metadata boosts
  -> regroup chunk matches into file-level results

Default storage:

~/.obsidian-semantic-mcp/
  data/semantic.sqlite

The vault remains the source of truth. The SQLite database is a derived index and can be deleted/rebuilt.

Safety Model

The server reads your vault and never writes to it. Three layers decide what an agent can see.

1. Always denied (system / tooling). Never indexed, no override:

• .obsidian/, .smart-env/, .claude/, .codex-*/
• any hidden folder (name starts with .)
• node_modules, cache, logs

2. Sensitive — denied by default, unlockable. Stays blocked even when a tool call passes include_sensitive: true, unless the server is started with OBSIDIAN_SEMANTIC_ALLOW_SENSITIVE=true. Defaults to 08_PersonalInfo/. Override the list with OBSIDIAN_SEMANTIC_SENSITIVE_PATHS (comma- or newline-separated folders):

OBSIDIAN_SEMANTIC_SENSITIVE_PATHS = "08_PersonalInfo, 09_Finance"

3. Your own excludes — always denied. Folders you never want indexed, searched, or read. No unlock flag:

OBSIDIAN_SEMANTIC_EXCLUDE = "03_Journal, Private, Clients/Acme"

Which one do you want?

• "Don't index this at all" → OBSIDIAN_SEMANTIC_EXCLUDE
• "Keep it locked, but I can unlock it with a flag when I need to" → OBSIDIAN_SEMANTIC_SENSITIVE_PATHS + OBSIDIAN_SEMANTIC_ALLOW_SENSITIVE

Additional guards:

• All paths are resolved through realpath.
• Path traversal and URL-encoded traversal are blocked.
• Symlinks that escape the vault root or point to a denied in-vault path are blocked.

search_notes and read_note enforce the current exclusion settings on every call. With automatic indexing enabled, the next search also removes excluded stale rows from the index. Confirm the active lists with index_status.

The local index stores snippets and embedding vectors. Treat it as a derived copy of your vault. See PRIVACY.md.

Indexing Strategy

The default is action-driven, not timer-based:

1. The first search_notes after server start.
2. The first search_notes after a Markdown create, update, delete, or rename event.
3. The first search_notes after the embedding model or index format changes; this triggers a full rebuild.
4. An explicit index_vault call.

The filesystem watcher only marks the index dirty. Indexing runs once, immediately before the next search, and concurrent searches share that work. If recursive watching is unavailable, the server falls back to an incremental scan before every search.

Disable automatic indexing with:

OBSIDIAN_SEMANTIC_AUTO_INDEX = "false"

Then call index_vault { "mode": "incremental" } yourself. The legacy OBSIDIAN_SEMANTIC_STARTUP_INDEX=true remains supported, but is not the default because it does work even when no search follows.

Search Quality Evaluation

Create a JSON file containing queries and expected paths to compare keyword, semantic, and hybrid modes under the same conditions:

[
  { "query": "strong consistency at 1000 TPS", "expected_paths": ["03_Knowledge/QnA/Strong-Consistency-at-1000TPS.md"] }
]

npm run evaluate -- ./queries.json 5

The command reports Recall@K and MRR. Avoid adding ranking complexity or embedding cost unless hybrid search measurably beats the keyword baseline.

Refresh the index with search_notes or index_vault before evaluating.

Development

npm test
npm run pack:check

Before publishing:

npm pack --dry-run

Confirm the package does not include data/, *.sqlite, or private vault files.

License

MIT