English | 한국어

🧠 Obsidian Everywhere

Turn linked notes into AI context, use that context from agents anywhere, and checkpoint approved changes with Git.

Graph context · local semantic search · remote agents · guarded edits · opt-in Git checkpoints

Watch the Remote Vault Bridge in 44 seconds

Remote setup guide

Remote request → semantic search → graph context → guarded edit → mount-loss recovery.

Obsidian Everywhere is built around three ideas:

1. Notes are a graph and a semantic knowledge base, not a folder of text files. Backlinks, n-hop neighborhoods, shortest paths, PageRank, full-text search, and local multilingual embeddings turn a topic into focused, token-budgeted context.

2. Your vault should be usable where your agents run. The Remote Vault Bridge exposes that same graph and its guarded write tools over authenticated Streamable HTTP. A Claude Code or Codex process on another server can search, reason over, append to, and reorganize a vault that remains on your own machine.

3. Version-control actions deserve a narrower boundary than file writes. If the vault or one configured folder inside it is already a Git repository, the opt-in Git tools can inspect status, bounded diffs, and local history. Commit and push each require a preview, explicit confirmation, and a short-lived one-use approval ID.

The local path stays the source of truth. Obsidian Everywhere does not create a hosted copy, telemetry service, or mandatory cloud account; optional Git push publishes only to a repository the operator already configured. Remote access is a transport you operate, not a vault-sync product.

Related MCP server: Vault Cortex Obsidian MCP Server

Contents

• Watch the Remote Vault Bridge in 44 seconds

• Features

• Three core workflows

• Vault Git

• Try it without your vault

• Why Obsidian Everywhere?

• Where does this actually run?

• Quickstart

• Configuration

• Development

• Project status

• Contributing

• License

Features

vault (.md files)
  │  parse · watch
  ▼
SQLite index (FTS5)  ⇄  in-memory graph (graphology)
  │                       n-hop · shortest path · PageRank
  ▼
41 core MCP tools + 0–5 opt-in Git tools
  │
  ▼
local stdio  ·  authenticated remote HTTP  ·  OAuth HTTP

• 🧩 Graph + semantic context engine — a markdown parser (wikilinks, embeds, frontmatter, nested tags, headings, block references), a SQLite index with full-text search, and an in-memory graphology layer for n-hop traversal, shortest paths, and PageRank. get_context_bundle packs a topic and its most useful neighbors into a requested token budget.

• 🌍 Remote Vault Bridge — agents on an external server get the same search, graph, context, and guarded editing tools over authenticated Streamable HTTP. Use a private network or an HTTPS tunnel such as ngrok; the vault itself stays on the machine you control.

• 📎 Vault-wide file reading — Markdown plus text/code/data files, PDF, DOCX, PPTX, XLSX, OpenDocument, EPUB, RTF, and common images are indexed and exposed without uploading them to a conversion service. Extraction is lazy, cached, size-limited, and searchable with search_files.

• 🧠 Optional local semantic search — semantic_search and get_related with method: "semantic" run a multilingual embedding model (multilingual-e5-small) entirely on your machine — no API key, cloud account, or Ollama process to run. Install the optional runtime with npm install @huggingface/transformers@^4.2.0; its model downloads once (~120MB, cached under ~/.obsidian-everywhere/), then works fully offline. It is disabled by default to keep the server below the 200 MiB memory target; opt in with OBSIDIAN_EVERYWHERE_ENABLE_SEMANTIC=true when extra memory is available.

• 🛡️ Safe writes and resilient mounts — partial edits, dry-run-first bulk operations, rollback snapshots, recoverable deletion, and an opt-in Beta mount guard. If a removable drive, NAS share, or container mount disappears, the index is preserved, writes are blocked, and a full reconciliation runs after it returns.

• 🌱 Reviewable Git checkpoints (off by default) — inspect status, bounded diffs, and local history for the vault or one configured repository folder. Higher modes add selected-file commits and operator-pinned HTTPS pushes, both behind preview and a five-minute one-use approval.

• 🛠️ 41 core MCP tools, up to 46 when Git is explicitly enabled — structured reads, attachment extraction, graph navigation, semantic retrieval, safe lifecycle operations, persisted Obsidian settings, and explicit health reporting.

Three core workflows

1. Turn a linked vault into focused AI context

Exact search finds the words you wrote. Semantic search finds the idea even when the wording or language differs. Graph traversal then explains how the matching notes relate. get_context_bundle combines those signals into a bounded context package instead of dumping an entire vault into the model.

2. Use and edit that context from an external server

Run Obsidian Everywhere beside the local vault, expose its HTTP endpoint through your private network or an HTTPS tunnel, and register the URL in the remote MCP client. The remote agent can read and search the local vault, then use the same guarded tools to create, append, move, tag, or clean up notes. The server reindexes each successful write before returning, so the next remote tool call sees the change.

For the complete ngrok path, see the Remote Vault Bridge with ngrok tutorial.

3. Review and checkpoint vault changes with Git

When the vault or one real folder inside it contains its own .git directory, an agent can inspect the same repository state you would inspect locally, then create a commit from an explicit file list. Set OBSIDIAN_EVERYWHERE_GIT_REPO_PATH to that vault-relative folder, or leave its default . to use the vault root. Push is a separate, stricter capability: it can only publish the current HEAD to its existing upstream branch through an operator-pinned HTTPS destination. It never pulls, fetches, changes branches, constructs a free-form refspec, or accepts arbitrary Git arguments.

Start with OBSIDIAN_EVERYWHERE_GIT_MODE=read; move to commit or push only after reviewing the safety model in the Vault Git guide.

Read

Write

Vault Git — registered only when explicitly enabled

If the ordinary write gate is disabled, git_commit and git_push stay absent even when the configured Git mode is higher; the three Git read tools remain available. OAuth therefore requires both a sufficient Git mode and OAUTH_ENABLE_WRITE_TOOLS=true for commit or push.

Ordinary write tools are on by default for stdio and the bearer-token HTTP transport, and off by default for the public OAuth connector transport (opt in with OAUTH_ENABLE_WRITE_TOOLS=true) — see Configuration and DECISIONS.md D15. Git is independently off by default on every transport.

Vault Git

Vault Git is an optional checkpoint-and-publish layer for repositories already inside a vault. It is not a sync engine and it never initializes a repository. Git must be installed on the vault machine. The operator selects exactly one repository with OBSIDIAN_EVERYWHERE_GIT_REPO_PATH, a safe vault-relative real directory that defaults to .. That selected directory must be the exact root of a normal repository with a real, local .git directory. The rest of the vault remains indexed and available to ordinary graph, search, and note tools.

For example, a vault at /Volumes/SanDisk/jwhong can keep full-vault context while Git tools operate only on /Volumes/SanDisk/jwhong/DSLab:

export OBSIDIAN_VAULT_PATH=/Volumes/SanDisk/jwhong
export OBSIDIAN_EVERYWHERE_GIT_REPO_PATH=DSLab
export OBSIDIAN_EVERYWHERE_GIT_MODE=read

Git tool path inputs and outputs are relative to DSLab in that setup; ordinary note and file tool paths remain relative to the vault root. Parent repository discovery, linked worktrees, submodule roots, symlinked repository paths, and unsafe external or symlinked object/ref/log/core metadata layouts—including alternate object stores—are refused by every Git tool. The canonical vault, selected repository, and .git directory identities are captured at startup and rechecked before every Git subprocess; replacing a directory or introducing a symlink fails closed until the operator verifies the mount and restarts the service. Commit and push additionally refuse detached branches, shallow history, sparse checkouts, per-worktree Git configuration, grafts/replacement refs, and in-progress history operations.

With the supplied Compose file, set OBSIDIAN_VAULT_HOST_PATH=/Volumes/SanDisk/jwhong and OBSIDIAN_EVERYWHERE_HTTP_GIT_REPO_PATH=DSLab for the bearer service. The OAuth service has its own independent OBSIDIAN_EVERYWHERE_OAUTH_GIT_REPO_PATH input; both service-specific repository paths default to ..

Choose the narrowest capability that covers your workflow:

Push mode also requires a comma-separated operator mapping from each allowed upstream remote name to one exact credential-free HTTPS destination. Continuing the DSLab example above:

export OBSIDIAN_EVERYWHERE_GIT_MODE=push
export OBSIDIAN_EVERYWHERE_GIT_REPO_PATH=DSLab
export OBSIDIAN_EVERYWHERE_GIT_ALLOWED_PUSH_REMOTES=origin=https://github.com/owner/repo.git

Use OBSIDIAN_EVERYWHERE_GIT_REPO_PATH=. instead only when the whole vault is the repository.

The selected branch must already track a normal branch on the mapped remote, and that remote's sole resolved push URL must exactly match the pinned mapping. The URL is operator configuration, never MCP tool input; credentials, queries, fragments, caller-selected branches, and caller-selected refspecs are refused. Git must authenticate to that exact URL non-interactively using credentials already configured on the vault machine.

The destination ref comes from the branch's existing upstream mapping, not from the local branch name. For example, local main tracking origin/release can push only to release; the caller cannot substitute another branch.

Commit and push are deliberately two-step operations:

git_status
git_diff
git_commit { action: "preview", message: "docs: update project notes", paths: ["Projects/Atlas.md"] }
  → inspect the plan and explicitly approve it
git_commit { action: "execute", approvalId: "<UUID from preview>" }

git_push { action: "preview" }
  → inspect the exact HEAD, upstream, and outgoing count; explicitly approve it
git_push { action: "execute", approvalId: "<UUID from preview>" }

An approval ID expires after five minutes, works once, and is invalidated when the reviewed repository state changes. A preview never creates a commit or contacts the network. Hidden, excluded, and sensitive paths are omitted or blocked; commits select exact changed paths whose resulting entries are regular files, plus deletions; hooks, signing, clean filters (including Git LFS), submodules, and suspected secrets are refused. Push review is capped at 100 outgoing commits and 200 changed blobs, with an 8 MiB per-file/blob and 32 MiB aggregate content limit; commit messages and merge results are scanned too. Commit messages are single-line and secret-scanned, and commit approval binds the exact proposed tree.

Push execution uses the displayed literal HTTPS destination and an exact OID lease for the reviewed upstream ref. That lease is a compare-and-swap guard—not permission for an arbitrary force-push—so a deleted, advanced, or reset remote ref fails instead of being overwritten. Repository-local credential helpers, URL rewrites, http.* transport settings, and selected-remote proxy overrides are also refused for push. Trusted HTTPS credentials and any required network policy belong in the vault machine's user or system Git configuration, outside the repository.

There is intentionally no git_exec or free-form command tool. Passing raw Git arguments to a remote agent is effectively a remote-code-execution primitive: Git aliases can expand to shell commands, hooks execute programs, diff/textconv drivers run helpers, SSH transports launch commands, and credential helpers may invoke executables. A small set of fixed commands with fixed arguments is the safety boundary, not a cosmetic API choice.

Read the complete setup, operational limits, and troubleshooting guide before enabling commit or push: Using a Git-backed vault.

Try it without your vault

Run the built-in demo first. It creates a temporary sample vault, shows graph orientation and unresolved-link discovery, previews a safe bulk edit, and then removes the sample. It never reads or changes your own notes.

npx -y obsidian-everywhere demo

When you are ready to connect a real vault, generate copyable configuration for Codex, ChatGPT Desktop, Claude Code, and Claude Desktop:

npx -y obsidian-everywhere init /absolute/path/to/your/vault
npx -y obsidian-everywhere doctor /absolute/path/to/your/vault

init only prints configuration—it never edits global client settings. doctor checks Node.js, permissions, Obsidian metadata, SQLite, parsing, and the graph engine without printing note content. Add --share to redact the vault path before pasting diagnostics into an issue.

Why Obsidian Everywhere?

There are several good Obsidian MCPs. Pick the architecture that matches how you work rather than assuming one server wins every category.

Comparison checked against each project's published documentation on 2026-07-20. A blank or narrower cell means “not documented there,” not that a project can never support it. If you need active-file state or command-palette execution, choose a plugin-backed server. If you want a headless, one-command graph server with token-budgeted context, guarded cleanup, and narrowly scoped Git checkpoints, that is the niche Obsidian Everywhere is designed for.

Everything runs locally by default. There is no account, API key, hosted vault, or telemetry requirement.

See docs/architecture.md for how it's built, docs/deploy.md for the deployment topology, and docs/ngrok-remote.md for an end-to-end external server tutorial. Git-backed vault operators should also read docs/git-vault.md.

Where does this actually run?

The obsidian-everywhere process needs direct filesystem access to your vault's .md files (to parse them, watch for changes, etc.) — so it must always run on the machine where your vault physically lives ("the vault machine": your laptop, most likely). It does not matter which client machine you're working from — the server always runs on the vault machine; only the client connection method changes.

You can run more than one of these at once (e.g. stdio on your laptop and bearer-token HTTP for your work server) — they're independent processes that all index the same vault.

Quickstart

The fastest install needs no clone or build step. Run this on the vault machine (wherever your .md files live):

npx -y obsidian-everywhere /absolute/path/to/your/vault

MCP clients normally launch this command for you using one of the configurations below.

Not sure whether the path and runtime are ready? Run the privacy-safe diagnostic:

npx -y obsidian-everywhere doctor /absolute/path/to/your/vault

Option A — Codex CLI and ChatGPT Desktop, same machine as the vault (stdio)

Codex CLI, the Codex IDE extension, and ChatGPT Desktop's Codex experience share the same MCP configuration (official MCP documentation). Add the server once:

codex mcp add obsidian-everywhere -- npx -y obsidian-everywhere /absolute/path/to/your/vault
codex mcp list

Then restart ChatGPT Desktop (or the IDE extension). In ChatGPT Desktop you can also add it through Settings → MCP servers → Add server, choose STDIO, and enter the same command and arguments. Type /mcp in Codex to confirm that the expected tools are connected: 41 with ordinary writes enabled and Git off, or the conditional counts documented in Vault Git.

For a project-scoped configuration instead, add this to a trusted project's .codex/config.toml; use ~/.codex/config.toml to make it available globally:

[mcp_servers.obsidian-everywhere]
command = "npx"
args = ["-y", "obsidian-everywhere", "/absolute/path/to/your/vault"]
startup_timeout_sec = 30

Use an absolute vault path. GUI apps may not inherit the same PATH as your terminal; if npx is not found, replace command with the absolute result of command -v npx.

Option A′ — Claude Code, same machine as the vault (stdio)

Still on the vault machine:

claude mcp add obsidian-everywhere -- npx -y obsidian-everywhere /path/to/your/vault

Or with environment variables instead of a positional arg:

OBSIDIAN_VAULT_PATH=/path/to/your/vault claude mcp add obsidian-everywhere -- npx -y obsidian-everywhere

Option A″ — Claude Desktop, same machine as the vault

Add to claude_desktop_config.json on the vault machine:

{
  "mcpServers": {
    "obsidian-everywhere": {
      "command": "npx",
      "args": ["-y", "obsidian-everywhere", "/absolute/path/to/your/vault"]
    }
  }
}

Option A‴ — Google Antigravity CLI (agy)

Add to your global Antigravity MCP configuration file (~/.gemini/config/mcp_config.json):

{
  "mcpServers": {
    "obsidian-everywhere": {
      "command": "npx",
      "args": ["-y", "obsidian-everywhere", "/absolute/path/to/your/vault"]
    }
  }
}

Option B — Codex, ChatGPT Desktop, or Claude on a different machine

Choose one secure route to the vault machine:

• Private network: use Tailscale and follow the steps below.

• Public HTTPS tunnel: use the read-only-first ngrok Remote Vault Bridge tutorial. Never expose the local plaintext HTTP port directly.

Step 1 — set up a private network between the two machines, if you chose Tailscale:

# on BOTH the vault machine and the MCP client machine
curl -fsSL https://tailscale.com/install.sh | sh   # or: brew install tailscale (macOS)
tailscale up                                        # opens a browser to log in / join your "tailnet"
tailscale status                                    # confirm both machines can see each other

Note the vault machine's Tailscale hostname/IP from tailscale status (something like my-macbook.tailnet-name.ts.net or 100.x.y.z).

Step 2 — start the server, on the vault machine:

openssl rand -hex 32
# Save that output in a password manager, then use the same value on both machines.
export OBSIDIAN_EVERYWHERE_TOKEN="<saved token>"
OBSIDIAN_VAULT_PATH=/path/to/vault \
  npx -y --package obsidian-everywhere obsidian-everywhere-http

Keep the saved token — you'll need it in step 3. (To keep this running persistently instead of in a foreground terminal, see the LaunchAgent setup in docs/deploy.md, or run it in Docker via docker-compose.yml if the vault machine is a server.)

Step 3 — connect from the other machine (the lab server, etc.), using the vault machine's Tailscale address from step 1. For Codex (and the shared ChatGPT Desktop configuration), keep the token in an environment variable:

export OBSIDIAN_EVERYWHERE_CLIENT_TOKEN="<the token from step 2>"
codex mcp add obsidian-everywhere \
  --url http://<vault-machine-tailscale-name>:3737/mcp \
  --bearer-token-env-var OBSIDIAN_EVERYWHERE_CLIENT_TOKEN

Ensure ChatGPT Desktop is launched with that environment variable available, then restart it. Alternatively, use Settings → MCP servers to add the Streamable HTTP URL and bearer credential if your app version exposes those fields.

For Claude Code:

claude mcp add --transport http obsidian-everywhere \
  http://<vault-machine-tailscale-name>:3737/mcp \
  --header "Authorization: Bearer <the token from step 2>"

The second machine now has access to the vault indexed on the first. Full walkthrough (Docker, LaunchAgent): docs/deploy.md.

Option C — claude.ai web/mobile app (custom connector, OAuth)

This needs a public HTTPS endpoint — claude.ai's servers can't reach your Tailscale network or localhost. See docs/deploy.md for the full Cloudflare Tunnel walkthrough. Once your server is reachable at https://your-domain:

1. claude.ai → Settings → Connectors → Add custom connector

2. Server URL: https://your-domain/mcp

3. claude.ai auto-discovers the OAuth flow and shows this server's sign-in page — enter the OAUTH_LOGIN_SECRET you configured.

You only need this if you actually want claude.ai's web/mobile apps to read your vault. If you only ever use Claude Code (locally or from another machine), skip this entirely — Option A/B already fully covers that with no Cloudflare/OAuth involved.

Configuration

Obsidian Everywhere does not modify a user's vault .gitignore. Built-in Vault Git rejects hidden .obsidian-everywhere paths, but if an index directory falls inside a repository you also manage with ordinary Git, add .obsidian-everywhere/ to that repository's own .gitignore.

Git configuration is independent of semantic search and ordinary note tools. For stdio and bearer HTTP, OBSIDIAN_EVERYWHERE_READONLY=true removes both ordinary write tools and Git commit/push. For OAuth, commit/push require both OBSIDIAN_EVERYWHERE_GIT_MODE=commit|push and OAUTH_ENABLE_WRITE_TOOLS=true. git_status, git_diff, and git_log remain read-only tools at every mode above off.

Direct CLI, HTTP, OAuth, and LaunchAgent processes read the generic Git names. The supplied docker-compose.yml deliberately uses the service-specific HTTP/OAuth .env inputs above so enabling Git for one exposed service cannot silently enable it for the other.

Development

npm run dev:stdio          # tsx, no build step
npm run dev:http
npm run dev:oauth-http
npm test                   # vitest, runs against fixtures/test-vault
npm run typecheck
npm run lint
npm run format:check
npm run memory:smoke      # asserts the default attachment workload stays below 200 MiB RSS

fixtures/test-vault/ is a 30+ note fixture vault exercising every link and parsing edge case the parser needs to handle (piped aliases, heading and block links, embeds, frontmatter-embedded wikilinks, nested tags, duplicate filenames across folders, unresolved links, code-block exclusion, and Korean filenames/tags/wikilinks). It's what every test in src/**/*.test.ts runs against.

Project status

The current release line includes the graph and optional local semantic context engine, all three transports (stdio, bearer HTTP, OAuth HTTP), 41 core MCP tools, guarded partial and bulk writes, and client setup for Codex, ChatGPT Desktop, and Claude. Remote Vault Bridge is a first-class deployment path. Its opt-in mount guard remains Beta while it receives cross-platform feedback for removable drives, NAS shares, and container mounts. The separately opt-in Vault Git tools expose read, commit, and push as progressively wider, review-gated capabilities instead of a general-purpose Git shell.

Help test a real remote-vault setup in Beta Issue #18, or ask questions in Discussion #19. A disposable vault is welcome; never share note contents, tokens, or private hostnames.

Contributing

Bug reports, feature requests, and PRs are welcome — see CONTRIBUTING.md for dev setup, testing conventions, and how the fixture vault relates to the test suite. Security issues: please see SECURITY.md rather than opening a public issue.

License

MIT — see LICENSE.