Seekstone

The Obsidian MCP server that needs no plugin, no running Obsidian app — and doesn't blow your context window. Filesystem-direct · single-digit-ms keyword search · ~26 ms semantic · ~2 KB payloads · 25 tools · macOS · Linux · Windows

seekstone.dev →

Same queries, same committed vaults, 20 runs each, one machine. In-process adapters were re-run on the fixture-v2 vault in September 2026; the REST-proxy rows (rest, obsidian-mcp-server, mcp-obsidian) plus obsidian-mcp and obsidian-mcp-pro are June 2026 captures on fixture v1 — per-row provenance in benchmarks.json, the generated source of truth every number here is checked against in CI — full results across eight servers and three vault sizes below, fully reproducible from the harness.

What is Seekstone?

Seekstone is an Obsidian MCP server — it gives Claude (and any Model Context Protocol client) direct read and write access to your Obsidian vault. No Obsidian app needs to be open, no plugins are required, and nothing leaves your machine.

It reads your vault directly from disk rather than routing through the Obsidian Local REST API plugin, and holds a warm full-text index in-process. The practical difference is twofold:

• Speed. Keyword searches return in single-digit milliseconds warm and semantic searches in ~26 ms — up to ~508× faster than every other Obsidian MCP server we benchmarked, because there's no subprocess to spawn and no HTTP round-trip per query.
• Context. A broad search that returns tens of megabytes and millions of tokens via a REST-proxy server returns ~2 KB via Seekstone — up to a ~47,000× reduction that only widens as your vault grows.

Search comes in three modes: ranked full-text search (fuzzy and prefix matching), optional local semantic search (meaning-based, via a small on-device embedding model — opt-in, offline at runtime after a one-time ~30 MB model download), and structured metadata queries — query_notes filters by frontmatter properties (status, due, type, …), tags, folder, modified time, and size, answering questions like *"which draft notes changed this week?"* in a few hundred bytes instead of a search-and-read loop.

Claude can search and read your entire note library, in milliseconds, without burning most of its context window on a single tool call.

Published on npm as seekstone — install with npx -y seekstone. (Previously also published as obsidian-mcp-seekstone; that alias is deprecated but existing installs keep working.)

Why Seekstone? The numbers.

Most Obsidian MCP servers return full note content for every search hit. On a broad query that's megabytes of text your LLM has to process — most of it irrelevant, all of it burning context window.

Seekstone returns short ranked excerpts instead (~120 characters by default, tunable per query). We benchmarked Seekstone against 7 other Obsidian MCP servers — 8 servers total — across three vault sizes — 1,000 / 5,000 / 10,000 notes (20 runs each). Every number below is fully reproducible: the vaults are committed to this repo (generated from the public-domain 1911 Encyclopædia Britannica), so you can clone it and run the exact same benchmark yourself.

The point of testing three sizes is that this is where the architectures diverge — a real vault only grows.

Search payload — bytes returned per query (context tax; lower is better)

Seekstone stays flat (~2 KB) no matter how big your vault gets, because it always returns ranked excerpts — and it's now the smallest payload of every server tested, edging out mcpvault at all three sizes. The REST-proxy servers return full note content for every match, so they grow with the vault — mcp-obsidian hits 95 MB at 10k notes, and a single broad query (the capital of) averaged 370.9 MB / 97.8 million tokens per call across 20 runs. At 10k notes that's a ~47,000× context-tax difference.

Search latency — warm mean, ms (lower is better)

Every competitor spawns a subprocess or makes HTTP round-trips per query, and most do work that scales with vault size. Seekstone holds a warm in-process index — no IPC, no network — so keyword search stays in single-digit milliseconds even at 10,000 notes (the shipped semantic pipeline — embed, scan, MaxSim rerank — lands at ~26 ms). And the gap widens with scale: from 1k → 10k notes the competitors slow down 5–10×, while Seekstone barely moves. Even the *fastest* alternative — obsidian-mcp-rs, which re-scans the vault on every query — is ~6× slower warm at 10k notes with 3× the payload, and the REST-proxy generation runs ~100–280× slower.

Tool list — tokens loaded before the first call (lower is better)

When a session starts, most MCP clients load the server's tool list (each tool's name, description, and input schema) into the model's context, before any tool is called. Some, such as Claude Code with tool search, load only the names until a tool is needed. Loaded in full, Seekstone's 25 tools cost ~5,742 tokens (26 KB). With SEEKSTONE_READ_ONLY=1 the 11 write tools drop out of the list and it falls to ~3,261 tokens, 43% less. The harness measures this from the built server over MCP (npm run harness -- schema-tax), and CI re-me