Vantell

Share what you know — never your notes.

Vantell lets colleagues discover what topics you have knowledge about, without your notes ever leaving your vault. You choose folders you're willing to share from; Vantell publishes only topic labels and note counts from them. Everything else — and all note contents, always — stays on your machine.

Install

In Obsidian, open Settings → Community plugins → Browse, search for “Vantell”, then Install and Enable.

The Vantell panel

Click the radio icon in the left ribbon (or run the “Open the panel” command) for a live side panel: your mesh as a conversation list, one row per colleague, showing the last thing that passed between you and a badge when someone is waiting on your answer. Open a row and the thread expands in place — their questions and answers on one side, yours on the other — with a composer at the foot for your next question. Requests carry their Approve / Decline buttons right in the thread, and a header toggle shows the mesh as a small constellation. It's the interactive home for the mesh — everything else (setup, per-note sharing) still lives in commands and settings.

Writing an answer still opens the answer window, where you pick which note titles to list as sources and can build a drafting prompt for your own Claude. Nothing is ever sent until you press Send.

When someone knocks

A waiting question puts a count on the Vantell ribbon icon and an in-app notice that stays until you deal with it — no eight-second toast you can miss by looking away. Obsidian also checks when you come back to the window, not only on its two-minute timer.

Optionally (Settings → When someone knocks) Vantell shows a system notification as well. It names who knocked and nothing else: never the question, never the reason they gave. Those are someone else's words, and a notification lands on your lock screen in front of whoever is standing there. Turning it on is when your OS is asked for permission; it is off until then, and it uses the browser notification API — no extra process, nothing new on the network.

How it works

1. Link your account — enter a one-time code from app.vantell.ai (free account, 2 minutes to create).
2. Choose your folders — everything starts private. Pick the folders you'd share from, and whether each is visible to your team or your whole org.
3. Review — see the exact list of topic labels and counts before anything is sent. What you see is literally all there is; the payload format has no field that could carry note contents.
4. Go live — colleagues can now find you by what you know and send you a request. You approve or decline each one right in the panel (or on your dashboard); every decision is recorded in a receipt you can review.

What leaves your vault — the complete list

All of this plugin's network requests go to api.vantell.ai (or the server you configured). The complete route list sits at the top of src/api.ts and is checked against the source by a lint, so it cannot silently drift. What is sent falls into the rows below, each shown to you before it happens:

When	What is sent
Linking (once per device)	Your one-time code + this device's public key
Going live / updating	Topic labels, note counts, the names of folders you chose, up to three sample note titles per chosen folder (displayed verbatim in the pre-publish review), and a signed timestamp
While Obsidian is open (every 2 min)	A signed check for incoming requests — an empty-bodied read of your own inbox and consent state. Nothing about your vault is sent
When you answer a request	Only the text you typed in the answer box, plus the titles you explicitly ticked as sources — after you approved the request (in the panel or on your dashboard)
If you turn on AI drafting (off by default)	The text of shareable notes matching the request's topic, sent to your own Anthropic account with your API key — never to Vantell, never to the person who asked. See below

Never sent: note contents (except AI drafting, below), note titles beyond the per-folder samples above and the ones you explicitly tick as answer sources, names of unchosen folders (they're reported only as one anonymous total), file paths, or any telemetry. Incoming requests are displayed as inert text and never written into your vault or executed.

Links the plugin opens (in your browser, with nothing attached): app.vantell.ai — your dashboard, for pairing codes and knock approvals — and claude.ai, an optional shortcut offered by the paste-bridge drafting path. These two domains appear in the code only as links; the plugin's network requests go exclusively to api.vantell.ai (or the server you configured).

Why the code contains atob/btoa: sealed envelope bytes (and keys) are base64-wrapped for transport — the encoding defined by the open Knock Protocol, not obfuscation. Envelopes are unsealed with this device's key and shown to you as inert text.

Answer drafting — two ways, both optional

When you answer a request you can draft it yourself, or get a hand:

Draft with your own Claude (no setup, any plan). The answer screen has a “Draft with my Claude” button. By default it builds a lean prompt: the question, the context, and the paths of the shareable notes most relevant to the topic — and lets your own Claude open and read them. This is ideal when your Claude can see the vault (Claude Code running in the folder, or the folder attached), and keeps the prompt tiny. If your Claude can't open the vault (e.g. plain claude.ai), click “Include note text” and it switches to a self-contained prompt with condensed excerpts. Either way you copy it, paste into your Claude, paste the answer back, and Send. The plugin sends nothing anywhere; the note text moves only when you paste it. Works on Free, Pro, Max, or Team — no API key.

Automatic drafting (advanced, needs an Anthropic API key). If you have a developer key from console.anthropic.com, add it in settings and a second button drafts in place. That path sends the same shareable, topic-matching notes to your own Anthropic account (never to Vantell, never to the person who asked); the key is stored on this device only and never synced. Most people don't need this — the paste path above covers every Claude plan.

Auto-drafting (optional, works on a normal Claude subscription). The full, honest explanation — what the helper can see, why the install needs an admin password, and the one-command undo — lives at app.vantell.ai/auto-drafting (no login needed) — in short: with standing consent you can let your second brain pre-draft answers for a specific person, so a draft is already waiting when you open the panel. This runs in a separate companion process under its own OS user that provably cannot read your vault — it only sees copies of notes you already cleared for that person's audience, and it re-proves that separation every time it starts (if the check fails, drafting turns itself off). Setup is three steps: install the companion (one admin command — see companion/ in this repo), run "Vantell: Set up auto-drafting" in Obsidian, and paste a credential in settings — on a normal Claude Pro/Max plan, run claude setup-token in a terminal and paste the result (an Anthropic API key in its own capped workspace is the stricter alternative). Per person it is capped at 10 drafts a day, revocable like any standing consent, and you still read and press Send on every answer — the plugin cannot send anything by itself.

Either way, the draft lands in the editable box, you review and edit it, and nothing reaches the person who asked until you click Send. Folders that look like they contain other people's words (emails, transcripts, chat exports) or notes about people are excluded automatically — even inside folders you chose — and this cannot be overridden.

Does Vantell change my notes?

No — with one deliberate, visible exception. Your folder choices are saved to a small settings file (.vantell.yml) in your vault, not to any note. The only time a note is touched is when you run "Share this note" or "Stop sharing this note" on that specific note: one visibility property is added, visible in Obsidian's Properties panel, removable at any time. The note's content is never altered (the plugin verifies the body survives byte-for-byte before writing).

Leaving is a first-class feature

"Remove Vantell from this vault" (in settings, or as a command) puts the vault back the way Vantell found it: it takes your published listing off the mesh (on by default — topic labels, counts and folder names disappear for colleagues), strips Vantell's sharing properties from every note (bodies verified byte-for-byte; pre-existing metadata like topics is never touched), deletes the .vantell.yml settings file, and deletes this device's signing key — after showing you the exact list of affected notes. The same "Remove published data" action also exists on your dashboard (Settings → Data & deletion), usable even after the device key is gone. Then uninstall the plugin like any other.

Security

• Your device signs its requests with a key generated locally (Ed25519). The key is stored device-locally and is never synced with your vault — not through Obsidian Sync, iCloud, or git — and never written into any note or plugin data file.
• Your mesh history — incoming questions, received answers, the knocks you sent — is stored device-locally too, never inside the vault folder. Vault sync, backups, git, and vault-reading tools (including your own AI) never see it; the plugin's data.json holds only preferences and your already-public listing summary.
• No password is ever entered in the plugin. Linking uses a single-use code that expires in 60 minutes.
• Knock and answer envelopes are end-to-end sealed (libsodium crypto_box_seal) to the recipient's key — the relay stores ciphertext it cannot read, and since 2026-08-14 it rejects unsealed envelopes outright. What the relay necessarily sees is routing metadata: who asked whom, when, at what level, and the stated purpose.
• Honest limitation: the signing key and the optional Anthropic key are stored device-locally, protected by your OS user account — not in an OS keychain. Obsidian plugins are not sandboxed from one another, so anything running as you (including other plugins) could read them. Both are deleted by "Remove Vantell from this vault".
• The plugin's full source — including the vault scanner (vaultscan-core) — is public: github.com/goffiti/vantell-obsidian. Audit it, or ask your own AI to.

Requirements & disclosures

• Requires a free Vantell account (your work email determines your org). The plugin depends on the Vantell service to relay what you publish; it does nothing without an account.
• Network use: only the calls listed above, only to your Vantell server. The scan itself makes zero network calls.
• No telemetry. The plugin collects no analytics of any kind.

For teams

Vantell implements the Knock Protocol: colleagues' requests to go deeper than topic labels arrive as knocks you explicitly approve or deny, with consent receipts and a content-free audit trail. Your org admin sees metadata only — never contents, never question texts — and envelopes are sealed end-to-end, so the relay operator can't read them either.