vault-mcp
Remote MCP server for Obsidian vaults, enabling secure access to markdown notes from Claude clients without the Obsidian desktop app. Provides tools for searching, reading, listing recent notes, resolving daily notes, and creating/appending notes via HTTPS.
README
vault-mcp
[!WARNING] Work in progress. The security model and core workflows are implemented and tested, but the project has not reached a stable release. Expect breaking configuration changes, review the threat model before deployment, and do not treat the current
mainbranch as production-ready yet.
A remote MCP server for your Obsidian vault. Add it once as a connector in Claude's settings and your notes become available in any conversation — on Claude web, iOS, Android or Desktop — with no Obsidian desktop app required anywhere. The server reads the markdown files directly from disk on a small VPS that stays in sync with your vault.
Your vault stays a folder of markdown files. No database, no proprietary index, no custom format. If this project disappears tomorrow, your notes are exactly where they were.
How it works
Three zones with strictly separated responsibilities:
Claude web / iOS / Android / Desktop
│
│ MCP over HTTPS (OAuth 2.1 bearer token)
▼
┌─────────────────────────────────────┐
│ EDGE — Cloudflare Worker │
│ · OAuth 2.1 + PKCE, consent page │
│ · token storage (Workers KV) │
│ · fixed redirect_uri allowlist │
│ · forwards with x-origin-secret │
│ (never stores vault content) │
└─────────────────────────────────────┘
│
│ Cloudflare Tunnel — outbound-only
│ from the VPS; no inbound ports
▼
┌─────────────────────────────────────┐
│ ORIGIN — your VPS │
│ · mcp-server, bound to loopback │
│ answers 404 without the secret │
│ · vault-core: path validation, │
│ atomic writes, search │
│ · vault-guards: read/write │
│ sanitization, warn heuristics │
│ · ~/vault ◀─ sync client ─▶ your │
│ sync service (bidirectional) │
│ · git autocommit every 30 min │
└─────────────────────────────────────┘
The client never reaches the VPS directly. The Worker knows nothing about the vault. The origin is not reachable from the internet — it only receives traffic through the outbound tunnel, and answers 404 to anything that does not carry the shared origin secret.
The sixteen tools
| Tool | Type | What it actually does |
|---|---|---|
search_notes |
read | Case-insensitive literal substring search across markdown notes (no regex, no semantic ranking). Returns path, line number and snippet. Supports pagination (offset), folder scoping (path_prefix), tag filtering (tag) and visiting newest notes first (sort_by: mtime). Low-trust folders (imported clippings) are excluded unless explicitly included. |
read_note |
read | Full content of one note by vault-relative path, with a header reporting size and the note's version hash (for expected_hash on later edits). Output is sanitized (see below); very large notes are truncated and flagged. |
read_notes |
read | Up to 10 notes in one call; per-note errors are reported inline. |
list_recent |
read | Most recently modified notes, newest first (paths and timestamps only). |
get_vault_tree |
read | Every folder with its note count — the vault's table of contents. Structure only, never content. |
get_daily_note |
read | Resolves the daily note for a date (default today) using the vault's own settings — core Daily Notes (.obsidian/daily-notes.json) or the Periodic Notes plugin. If the note doesn't exist it returns the path it would have — it never creates it. |
create_daily_note |
write | Creates the daily note at the configured location, seeded from the configured daily-notes template ({{title}}, {{date}}, {{time}}, {{date:FORMAT}}). Idempotent: an existing note is left untouched. |
create_note |
write | Creates a new note. Fails if the note already exists — it never overwrites. Parent folders are created as needed. Writes are atomic. |
append_to_note |
write | Appends to the end of an existing note — the note must already exist (create it first), and existing content is never edited or overwritten. |
append_to_section |
write | Inserts at the end of a specific heading's section (before the next same-or-higher-level heading; code fences don't count as headings) — capture into ## 📥 Inbox without landing after a trailing dataview block. |
edit_note |
write | Exact search-and-replace edits, atomic and all-or-nothing: each old_string must match exactly once; supports expected_hash, checked again immediately before replacement, so stale edits normally fail with CONFLICT. Plain filesystems provide no portable CAS against unrelated external writers, so a narrow final race remains. |
move_note |
write | Moves/renames a note. Never overwrites the destination. Wiki-links are not rewritten. |
delete_note |
write | Moves the note to the vault's own .trash/ (same as Obsidian's "move to vault trash") — nothing is permanently erased. |
list_tasks |
read | Checkbox tasks across the vault, parsed with Obsidian Tasks plugin conventions (📅 ⏳ 🛫 ✅, priorities, 🔁). Filters by status, due-date window and folder; sorted by due date; each task reports path, line and note version hash. |
complete_task |
write | Flips [ ] to [x] and appends ✅ YYYY-MM-DD in the exact Tasks-plugin format. Recurring tasks are completed but the next occurrence is not generated. |
postpone_task |
write | Changes (or sets) a task's 📅 due date in the exact Tasks-plugin format. |
Remote images in written content are de-embedded into plain links before touching disk, and note content returned to the model is stripped of channels for invisible instructions (HTML comments, CSS-hidden elements, invisible characters, the Unicode tag block). See the threat model for why.
Non-goals — on purpose
- Single user, single vault. One instance serves one person. Multi-tenancy is a declared non-goal: it reintroduces an entire class of isolation problems that simply doesn't exist today. Each user runs their own instance.
- No Obsidian runtime. No Templater, no Dataview, no plugins, no
eval. Notes created through the server are plain markdown; plugin syntax in your templates will not be expanded. - No
delete_note, nomove_note. Deliberately absent: they turn noise into data loss, and a bad move breaks wikilinks across every synced device. The minimal tool inventory is a security control, not an oversight. - No HTTP-request tools, no shell, no JavaScript execution. Closing these channels is what keeps the worst case of a successful prompt injection at "junk in a note" instead of exfiltration.
- No semantic index / embeddings in this version. Full-text search covers most cases and adds no state to maintain.
Getting started
Follow docs/setup.md end to end: VPS bootstrap, sync client, tunnel, Worker deploy, and adding the connector in Claude's settings. There is also a local-only mode for trying the server on your own machine without any of the edge pieces.
Before hosting this, read docs/threat-model.md — you are exposing personal notes to the internet, even behind authentication, and you should understand exactly what protects them and what the residual risks are. Day-2 procedures (token revocation, secret rotation, restore from git) live in docs/operations.md.
Repository layout
apps/
mcp-server/ MCP server on the VPS; executes the tools over HTTP
auth-worker/ Cloudflare Worker at the edge; OAuth + authenticated proxy
packages/
vault-core/ reading, atomic writes, path validation, search
vault-guards/ input/output sanitization, warn-only heuristics
tool-contract/ tool schemas and descriptions, shared
infra/
scripts/ bootstrap · configure · start · doctor · autocommit
systemd/ user units for server, sync, tunnel, autocommit
tunnel/ cloudflared configuration example
docs/
setup.md step-by-step installation
threat-model.md assets, adversaries, defenses, residual risk
operations.md runbooks: revoke, rotate, restore
The architecture rationale (in Portuguese) is in ARCHITECTURE.md. The exact tool surface — names, schemas and the descriptions the model sees — lives in packages/tool-contract/src/index.ts.
Contributing
See CONTRIBUTING.md — including the dependency policy (a server with access to personal notes is a supply-chain target) and the standing answer to multi-tenancy requests.
License
MIT © Juliano Sirtori
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.