PageVault

Publish an HTML or Markdown artifact to a URL, and decide who can open it — straight from Claude, ChatGPT, Gemini, Copilot, or any MCP-capable tool, without leaving the chat where you made it. Self-hosted on Cloudflare's free tier. Whoever you send it to installs nothing and signs up for nothing — no account, on any platform.

   

This README is the setup guide. The product argument lives on the product page, and there's a live showcase running on PageVault itself if you'd rather just look at the thing.

---

Is this for you?

Two tiers, and picking one is really a single question: does anyone need to be stopped at the door?

• Just handing someone a report you made in a chat — a proposal, an analysis, a one-off? You want Public: a URL anyone you send it to can open, no login, nothing to install. Optionally on your own domain.
• Sharing your own work with named people — a document only the CFO and CEO should open? You want Secured: access gated to specific email addresses.
• Running a practice with recurring clients? Still Secured, plus portals — one durable URL per client, permissions on the client instead of the document, and an agent that searches the collection back months later.

What you get	Public	Secured
Share a /p/ URL anyone can open	✅	✅
Your own domain	optional	✅
Documents gated to named emails	—	✅
Per-client portals + agent read-back	—	✅
Cost	free · no card	a domain · a card on file

The one real jump is Public → Secured. Secured runs on Cloudflare Zero Trust, and Cloudflare wants a card on file before it will turn that on — the free plan is free and you won't be billed, but that's the seam, and pretending otherwise would waste your time. Portals on top of it are just a data model. Start Public; climb when you actually need the door. Your documents carry across untouched, and Public undoes cleanly.

---

1 · Public

Public links anyone with the URL can open — on workers.dev, or your own domain. No Zero Trust, no card. ~10 minutes.

You need: a Cloudflare account · Node 22+ · a Cloudflare API token · macOS, Linux, or Windows (which OS?). Full detail in docs/setup/prerequisites.md.

Create your Cloudflare API token

Either path below reaches your account with an API token — explicit, so a deploy can never land in the wrong account. Create it once, with room to climb.

Cloudflare → API Tokens → Create Custom Token, name it pagevault. Grant all of these now, so climbing the ladder never means re-scoping:

Type	Permission	Access	Needed for
Account	Workers Scripts	Edit	Public · deploy
Account	Workers KV Storage	Edit	Public · documents
Account	Account Settings	Read	Public · identify the account
Zone	Workers Routes	Edit	domain · custom domain
Zone	DNS	Edit	domain · the domain record
Account	Access: Apps and Policies	Edit	Secured · gated access
Account	Access: Organizations, Identity Providers, and Groups	Edit	Secured · the viewer group lives here — easy to miss

Install — pick a path

Three ways to stand PageVault up. They end in the same place; the setup steps here pick up from a deployed PageVault.

	Pick this when	How
npm (recommended)	you just want it running	npm install -g pagevault && pagevault init
git clone	you want to read or change the code	git clone, then make setup && make deploy && make verify
Hand it to your LLM	you'd rather be walked through it	paste your assistant the setup runbook and follow along

The npm path, in full:

npm install -g pagevault
pagevault init          # pastes your token, picks a tier, provisions and deploys — no clone

pagevault init walks you through the token, the tier, and your account, ships the Worker to Cloudflare, remembers where it landed in ~/.pagevault/, and points the CLI at it — so publishing works with no extra step. Later, pagevault upgrade redeploys after npm update -g pagevault.

git clone runs the same engine from source (make calls the same code the CLI does), and it's the path both tiers here are also tested on. Hand it to your LLM isn't a third mechanism: the runbook picks npm or clone and walks you through that one — give your assistant this URL and answer its questions:

https://raw.githubusercontent.com/danjamk/pagevault/main/docs/setup/ai-guided-setup.md

Publish something

pagevault publish report.html --public

init already pointed the CLI at your deployment, so there's nothing to configure first. Or publish straight from the conversation where you made the artifact — see Connect an agent.

/v/ and /admin fail closed at this tier. That's correct: you aren't using them yet. When you want them, climb.

---

2 · Secured

Your own domain, and documents only named people can open. They get a six-digit code by email — no account, no password, nothing installed.

Adds: a domain in the same Cloudflare account · Cloudflare Zero Trust enabled (a card on file; nothing is charged) · a second, narrow runtime token.

The domain and the gating are separate steps. A domain alone keeps you Public, on your own hostname; turning on Zero Trust is what makes you Secured. Re-run setup to change either — it shows your current choices and asks only for what's new:

pagevault init          # re-run: choose Secured, give it your hostname, redeploy

Once Access is on:

pagevault publish report.html --emails cfo@acme.com,ceo@acme.com

That grant is additive — it can admit a viewer, never silently revoke one — and it invents no portal for two people.

The two tokens, and why

The runtime token lives inside the Worker as its CF_API_TOKEN secret and keeps the viewer group in sync as you add and remove people. It is scoped to a single permission on purpose: a compromised Worker can edit one Access group and nothing else — never your KV, never your other Workers. See ADR-002.

Running a practice — portals

Permissions move from the document to the client. A portal is one durable URL per client; every artifact lands there, gated to their people. Adding someone to a client's team is one write, not fourteen.

pagevault portal-create acme --name "Acme Corp" --kind restricted
pagevault publish q3-review.html --portal acme
pagevault share acme cfo@acme.com          # and --remove when they leave
pagevault search acme "migration decision"

That last line is a search across everything you've ever sent Acme — and it's the same search your agent runs, which is the next section.

Portals are invisible until you ask for them. Nothing above this section required knowing what one is.

---

3 · Connect an agent

This is the part that isn't a link-sharer. The MCP server runs inside your Worker — nothing extra to host, no second service, no bridge process on your laptop.

Two things happen once it's connected.

You publish from the conversation that made the artifact. No exporting a file, finding it in Downloads, and uploading it somewhere. The chat that produced the report hands it to your PageVault and gives you back the URL.

And the collection reads back. search_portal and read_document mean that six months into an engagement, "what did we decide about the migration?" is a question your agent can answer out of the client's portal — over the same connection you published through. Publishing and remembering become the same act. This is why the server is remote rather than stdio: a stdio server can't run in a browser or on a phone, which is where artifacts actually get made (ADR-006).

Surface	How it authenticates
claude.ai · Desktop · mobile	OAuth 2.1 — you sign in with your own Cloudflare Access identity, nothing to paste. Needs Secured
Claude Code	a bearer token, the one the CLI already uses. Works at either tier
Hosted surfaces on a Public deployment	bridge with mcp-remote and the bearer — with no Access there's nothing for OAuth discovery to find

Same server, same thirteen tools, every surface. The tools carry annotations, so a host knows which are safe to auto-run and which should ask first — publishing over an existing document and minting a public link both require an explicit confirmation.

Full walkthrough: docs/setup/connect-mcp.md.

---

How it compares

The honest version, including the rows where the alternatives win. The full field guide puts ten tools across six capability groups on an interactive map — and it's served through PageVault, so it doubles as a live sample.

The short table

One representative per category, not ten names.

	PageVault	sharehtml	Link tool	Portal SaaS
Per-client collection — the unit is the client, not the link	✅	❌	❌	✅
Your AI agent can search it and read it back — ask a question months later, get the passage	✅	❌	❌	~ rare
Publish from a chat (MCP)	✅	❌	~ some	❌
Renders HTML and Markdown, hands back the source and a PDF	✅	~ HTML	~	❌ file host
Free, self-hosted, your own domain	✅	✅	❌	❌
Your infrastructure — no vendor holds the data	✅	✅	❌	❌
Who opened it, and when	~ Secured, 90 days	❌	~ opens	✅
Collaboration — comments, presence	❌	✅	❌	✅
Watermarks, NDA gate, named audit trail	❌	❌	❌	✅

No single row is unique. Plenty of tools render HTML, several publish from a chat, and portal suites have held collections for years. What's rare is the combination: free, self-hosted, renders the artifact, holds the collection, and lets an agent read it back. That's the claim — not any one row.

Use something else if…

If you need	Go to
Comments, reactions, live collaboration on the document	jonesphillip/sharehtml — same Cloudflare design, plus collaboration. PageVault has none of it, on purpose.
The host to never see your plaintext	An end-to-end-encrypted tool. PageVault's Worker can read what it serves; that's an honestly weaker guarantee.
Per-viewer watermarks, an NDA click-through, an audit trail you can hand a lawyer	A deal-room product. PageVault records who opened what and when, for three months — useful, but not evidence.
CRM, invoicing, e-sign around the client relationship	A client-portal SaaS. A much larger product; PageVault is deliberately none of it.
One link, right now, zero setup	Claude's own Publish button, or any quick link-sharer.
Fewer than ~5 clients	A shared folder per client. Said plainly because it's true.

---

Docs

I want to…	Start here
See what it does, without installing it	The feature tour — served through PageVault, so it's also a sample
Understand the design	docs/architecture.md, then the ADRs
Work through setup properly	docs/setup/prerequisites.md
Have my assistant set it up	docs/setup/ai-guided-setup.md
Look up a CLI command	docs/setup/cli-reference.md
Connect Claude to it	docs/setup/connect-mcp.md
Back it up	docs/setup/backup-and-restore.md
See how it was built with an agent	docs/engineering/how-i-built-this.md
Contribute	CONTRIBUTING.md

Everything is indexed in docs/README.md. Release history: CHANGELOG.md.

---

_{MIT · built by @danjamk · PageVault owes jonesphillip/sharehtml three ideas — the Access-provisioning script, the capability-token model, and the sandboxed iframe. They got there first.}