The #1 Obsidian MCP · freshness-aware

Your vault.
Every agent.
Fresh, cited memory.

Hybrid search covers Markdown and PDFs/OCR; structured tools parse Canvas, run Dataview-style LIST/TABLE queries, and execute supported Base filters. Read-only by default, with source and freshness evidence your agent can audit.

Install and connect See the evidence

Local-first Read-only by default enquire outbound during serve: 0

YOUR KNOWLEDGE DATABASE Obsidian vault

LOCAL MARKDOWN + PDF RETRIEVAL enquire-mcp BM25TF-IDFMLRRFBGEHNSW/int8

Claude Cursor ChatGPT Codex OpenClaw

Grounded answer cited source

“You worked on Apollo and logged an idea about velocity.”

Source: 99_Daily/2026-05-02.md · deterministic synthetic vault

46MCP tools

19MCP prompts

2,228public tests

13release gates

50+embedder languages

Built for knowledge that compounds

Not a file wrapper.
The complete intelligence stack.

Literal search is useful. Vector search is useful. Real long-term memory needs both — plus reranking, citations, freshness, graph context, document coverage, and production controls.

Freshness-aware recall beyond keyword overlap

Local Markdown/PDF retrieval combines BM25 + TF-IDF + multilingual ML embeddings + RRF, with local BGE reranking and HNSW/int8 vector search.

One memory. Every agent.

MCP-native access keeps the same knowledge available when you change model, client, or workflow.

Answers you can audit

Depending on result type, tools expose source paths and line spans, PDF pages, age_days/stale freshness signals, and per-signal scores.

research/RAG.md ¶ 44–51

Your files stay yours

enquire has no telemetry and initiates zero outbound HTTP during serve. Requested context is returned to the connected MCP client; cloud clients and tunnels remain a separate privacy boundary.

enquire outbound during serve: 0

Hybrid recall plus Obsidian-native tools

Hybrid retrieval covers Markdown and PDFs/OCR. Structured tools parse Canvas, run Dataview-style LIST/TABLE queries, and execute supported Base filters.

MDPDFBASECANVAS

The complete stack wins

One package.
No missing layers.

The comparison uses complete product combinations—not isolated features. A cross means the full row was not documented on the reviewed public surface.

Complete leadership standard	enquire-mcp	Smart Connections	Hybrid Search	File-wrapper MCP
MCP memory shared by every agent	✓	×	✓	✓
BM25 + TF-IDF + ML + RRF + BGE + HNSW/int8	✓	×	×	×
HyDE + bounded multi-query + context packs	✓	×	×	×
Cited paths/PDF pages + age_days/stale freshness	✓	×	×	×
Markdown/PDF hybrid + Canvas, Dataview, Base-filter tools	✓	×	×	×
Per-signal scores and stage explanations	✓	×	×	×
Read-only default + write gate + privacy filters	✓	×	×	×
Signed release + 9 gates + semver contract	✓	×	×	×

Reviewed 2026-07-30. ✓ = complete row built in; × = complete combination not documented. Partial sub-features may exist.

Review pinned evidence

Measured, reproducible, inspectable

Marketing that ships with receipts.

Concrete claims map to source, CI invariants, or disclosed measurements.

BGE RERANKER · NDCG@10 POINTS +15.5

Measured over plain hybrid retrieval in the public 60-query within-project ablation.

Read the methodology

BGE RERANKER · MRR POINTS +24.7

The same reproducible corpus, query set, relevance judgments, and comparison protocol.

Inspect the artifacts

RELEASE INTEGRITY 9/9

Lint, Node 22/24 tests, smoke, audit, coverage, version sync, docs, and outside-in drift checks.

View live CI

Signed supply chain. npm releases carry Sigstore provenance bound to the tagged main commit; stable and RC channels stay separate.

Verify on npm

One backend, three high-value workflows

Built for people who
need context to survive sessions.

Start with live scan and no index. Add hybrid retrieval, reranking, HNSW, PDFs, and OCR only when your vault needs them.

01 · BUILDERS

Coding agents that remember decisions

Give Claude Code, Cursor, and Codex durable access to architecture notes, incidents, contracts, and project history.

02 · RESEARCHERS

Evidence across notes and PDFs

Search across languages, cite the exact page, and preserve the original source instead of an extracted cloud summary.

03 · KNOWLEDGE WORKERS

A second brain every model can use

Meetings, journals, people, and decisions remain portable Markdown while every MCP client gets the same memory layer.

No account. No API key. No migration.

Start with live scan.
Add depth when you need it.

1
Install
One npm package. Node 22.13+.
2
Point at your vault
Keep Markdown where it already lives.
3
Connect an agent
Generate the strongest verified action for its client.

Basic live scan requires no indexing or model download.

Quick install build 4.0.0-rc.4

Install and connect Claude Code in two commands:

$ npm install -g @oomkapwn/enquire-mcp
$ claude mcp add obsidian -- enquire-mcp serve --vault /absolute/path/to/vault

Generate VS Code’s official native review-and-install URI:

$ npm install -g @oomkapwn/enquire-mcp
$ enquire-mcp configure --client vscode --vault /absolute/path/to/vault

Generate the exact codex mcp add command and TOML fallback:

$ npm install -g @oomkapwn/enquire-mcp
$ enquire-mcp configure --client codex --vault /absolute/path/to/vault

Print reviewed config for Claude Desktop, Cursor, Windsurf, HTTP, or every client:

$ npm install -g @oomkapwn/enquire-mcp
$ enquire-mcp configure --vault /absolute/path/to/vault

5-minute quickstart API reference Client templates

A first-class page for humans and agents

Give your AI the project in one fetch.

Structured metadata, crawlable proof, a compact AI index, and generated API docs make the product legible before an agent ever clones the repository.

AI INDEXllms.txt DEEP CONTEXTllms-ctx.txt CONTRIBUTOR MAPAGENTS.md GENERATED CONTRACTTypeDoc API

The practical questions

Before you give an agent
your second brain.

Do I need Obsidian installed?

No. enquire-mcp reads .md, .canvas, and .pdf files directly from disk and works against any Obsidian-format vault — the Obsidian desktop app does not need to be running.

Will it write to my vault?

Not unless you pass --enable-write. All 7 write tools are gated behind that flag, and the destructive ones support a dry_run preview.

Is my data sent anywhere?

enquire itself sends no telemetry and makes zero outbound HTTP calls during serve. It returns requested vault context to the MCP client you connect; cloud clients may process that context under their own privacy policy. setup, build-embeddings, and install-model may explicitly download ONNX weights from Hugging Face; a hybrid-tier first-run --apply orchestrates those same acquisitions, while install-ocr-lang downloads a Tesseract language pack.

What is the query performance?

Performance depends on vault size, hardware, model, and enabled retrieval layers. The public evidence includes a production report of 50–100ms BM25 top-10 at 1,771 chunks / 368 files and a reproducible synthetic benchmark showing 37–103x FTS5 speedup over linear scan at 100–1,000 notes. Run the built-in evaluation on your own vault before setting a latency SLO.

What languages are supported?

The default paraphrase-multilingual-MiniLM-L12-v2 embedder covers 50+ languages and was validated end-to-end on bilingual Russian + English vaults. The default cross-encoder reranker is rerank-bge (English-only; the only catalog alias verified end-to-end); multilingual reranker aliases currently fail their transformers.js tokenizer compatibility check. CJK/Thai/Khmer tokenization uses Intl.Segmenter.

Can I run it remotely?

Yes — `serve-http` exposes the same server over Streamable HTTP. Front it with Tailscale Funnel or Cloudflare Tunnel for HTTPS. It works with claude.ai web, ChatGPT custom GPTs, Cursor HTTP mode, and mobile MCP clients.

Build memory once. Use it from every agent.

Your knowledge should outlive
the chat window.

Install enquire-mcp Star on GitHub

Your vault.Every agent.Fresh, cited memory.

Not a file wrapper.The complete intelligence stack.