Vnsh

[vnsh logo]

vnsh

One workspace all your AI agents can read and write

[Build Status] [npm vnsh] [npm vnsh-mcp] [GitHub Action] [License]

Website • Get Started • Links • How It Works • API • Self-Hosting

[A vnsh workspace: an incident report shared as one link, rendered from markdown, editable by any agent holding it]

Right now you paste the same context into Claude Code, then Cursor, then Slack.

vnsh gives that context one address instead. Drop it once and get a link; every agent and every person you hand it to opens the same living document, and can change it. Encrypted in your browser before upload, so vnsh never sees it, and deleted 24 hours after the last edit.

kubectl logs pod/app | vn# https://vnsh.dev/w/k2p9xf...#w=...   edit link — read and write# https://vnsh.dev/w/k2p9xf...#r=...   view-only — read, never write

Why this and not a pastebin

A pastebin gives you a snapshot. A workspace has a stable address and a version, so the next agent writes back to the same place instead of starting a new one. Two properties make that safe to share:

• The server cannot read it. Content is encrypted client-side and the key rides in the URL fragment, which HTTP never transmits. What the server stores is ciphertext and a SHA-256 of a write token — not the token. The file's name is encrypted too, under the same key, so your recipient downloads analysis.py while vnsh holds a string it cannot decode. That is checkable from outside: forge a token and you get a 403.
• Two agents cannot silently clobber each other. Writes are conditional on the version you read. An unconditional write is refused outright; a stale one gets a 412 telling you to re-read and merge. Measured in production with five concurrent writers: one succeeded, four were told, nothing was lost.

Get started

For an agent — the point of the thing

Paste this into Claude Code, Cursor, OpenHands, Cline, Windsurf, Zed — anything that speaks MCP:

Set up vnsh workspaces — one link to hand work between people and agents: https://vnsh.dev/llms.txt

It reads the protocol, installs the MCP server, and writes a standing rule into its own instruction file so it keeps using workspaces afterwards. By hand instead:

claude mcp add vnsh -- npx -y [email protected]

The server exposes vnsh_artifact_create, vnsh_artifact_list, vnsh_artifact_read, vnsh_artifact_update, vnsh_workspace_create, vnsh_workspace_read, vnsh_workspace_update, vnsh_workspace_history, vnsh_workspace_restore, vnsh_workspace_renew and vnsh_workspace_open, plus vnsh_share, vnsh_share_file and vnsh_read for one-shot content.

From the terminal

npx vnsh                      # or: npm i -g vnsh                              # or: curl -sL vnsh.dev/i | sh   (one-shot links only)
vn ./report.html              # create a workspace from a filegit diff | vn                 # or from stdinvn read "<any vnsh url>"      # read one back — workspace, public, or legacy blobvn write "<edit url>" ./new   # replace the contents; refuses to clobbervn history "<workspace url>"  # list the latest 20 retained versionsvn restore "<edit url>" 3     # restore v3 as a new latest versionvn init .                     # teach agents in this project to use vnsh

vn init adds an idempotent managed section to the project's agent instruction file. It works across agent vendors and records only an anonymous "project initialized" dimension when that CLI later creates a workspace, so the experiment can be measured without collecting a project name or path.

--public publishes it unencrypted (see below). --blob explicitly makes a one-shot link. --ttl works for workspaces too, up to 168 hours; subsequent writes preserve that lifetime, and vn renew can extend it without an edit.

The dependency-free shell function from curl -sL vnsh.dev/i | sh handles one-shot links only. Workspaces need HKDF-SHA256 and AES-256-GCM, and the openssl that ships with macOS is LibreSSL, which has neither — it hands /w/ links to npx vnsh when Node is available and says so plainly when it is not.

From a browser

https://vnsh.dev — drop a file or paste, and get both links back. Nothing is uploaded before it is encrypted.

From CI

upload-to-vnsh uploads build output or failing test logs and prints a link in the job summary.

From Chrome

The extension previews vnsh links inline on GitHub, Slack and Discord, and bundles a screenshot, console errors and the page URL into one link with ⌘D.

The three kinds of link

The view-only tier is not a setting the server enforces — it is arithmetic. The content key is HKDF(secret, "vnsh/enc/v2"), a one-way derivation, so its holder can decrypt every version while being unable to recover the secret and therefore unable to derive a write token.

Public workspaces exist because a person opening a link has a browser doing the decryption for them, and an agent's fetch does not. A public workspace is stored as written and served as an ordinary document, so anything that speaks HTTP can read it with no key and no setup. The trade is stated where you choose it: vnsh can read a public workspace. It is never the default and never inferred, its visibility is fixed at creation, and changing it still requires the write token.

They are served from vnshcontent.dev, a separate registrable domain, and that is not cosmetic. A public document is written by a stranger and rendered as a top-level page, while reputation systems — Safe Browsing, mail gateways, corporate proxies — list a domain rather than a path. One abusive page on vnsh.dev would take the API, the site and every installed CLI, MCP server and extension down with it. A subdomain would not help; the unit is the registrable domain. Sandboxing is a separate matter and already handled: a public document is served with a sandbox CSP directive, so it loads into an opaque origin with no cookies, no storage, no network and no access to any other page. /w/ links stay on vnsh.dev because the key in the fragment is a real gate — the server has never seen their plaintext and neither can a crawler.

Don't assemble a public URL yourself. Creating one returns the exact link in the response's url field, which is also what keeps a self-hosted single-domain instance working.

How it works

you ──encrypt──▶ [ vnsh: ciphertext, no key ] ──decrypt──▶ agent / person                              │                    deleted 24h after the last write

Key schedule

S = random(32)                        root secret, lives only in the fragmentK = HKDF-SHA256(S, "vnsh/enc/v2")     content key — AES-256-GCMW = HKDF-SHA256(S, "vnsh/write/v2")   write token, sent as 64 hex charsH = SHA-256(W)                        the only derived value the server stores

Workspaces use AES-256-GCM, not the AES-256-CBC of one-shot blobs, because mutable content needs integrity: without an authentication tag, anyone able to rewrite storage — the host included — could flip ciphertext bits undetectably, which hollows out the whole guarantee. Nonces are random per write and prepended, never derived from a version number.

The viewer renders HTML and markdown in a frame with sandbox="allow-scripts" and deliberately no allow-same-origin, so content runs in an opaque origin and cannot read the key out of location.hash; an injected default-src 'none' removes its network access.

The whole protocol is specified in /llms.txt, creation included, so you can implement it in any language with a crypto library and run no vnsh code at all. Someone did, in about 200 lines.

API

Writes answer 428 without If-Match, 412 on a stale version, and 403 on a bad write token. Full reference in docs/api.md.

Security model

Accounts and permanent artifacts

Anonymous sharing remains account-free and temporary. Sign in at https://account.vnsh.dev with a magic link to keep newly created workspaces and /artifact/ pages until you delete them. Browser creates use the signed-in session automatically. For the CLI, run vn login and approve the device in your browser. MCP and CI can use an account token in VNSH_TOKEN; content is still encrypted locally for capability-link workspaces, and the account database stores only their ownership metadata, not keys or plaintext. During the free preview, each account can keep 100 documents and 1 GB total, including retained versions. Keep the returned link: its fragment is the only copy of the decryption/editing secret, so the account can manage retention and deletion but cannot recover a lost key.

Phase 1 account Artifacts add a separate collaboration mode: authenticated humans and Agent tokens can discover and version private Artifact content without carrying a fragment key. That convenience has a different security boundary—vnsh can technically read account Artifact content. It is never silently substituted for an anonymous encrypted workspace, and only a human may approve or publish by default.

vn loginvn artifact create report.md --title "Release report" --type reportvn artifact list --query release --status draftvn artifact read <artifact-id>vn artifact update <artifact-id> report.md --base-version 1 --change-summary "Verified production"

Add --json to any Artifact command when another program or Agent will consume the result.

What holds. vnsh cannot read encrypted content, cannot write to it, and cannot recover a write token from anything it stores. Rendered content runs with no same-origin access and no network, so a hostile document can neither reach the key nor send anything anywhere.

What does not.

• Handing someone a link hands them the key, including that agent's model provider. The 24-hour clock is what bounds this, not the encryption — and each write restarts it, so a workspace edited daily stays alive.
• The boundary is the client, not the transport. Whatever encrypts holds your plaintext first, and the MCP server and CLI both do. npx -y refetches the latest published version on every start; pin it ([email protected]), install it globally once, or build from source if you review what you run.
• A public workspace is readable by vnsh, by design. That is the tier.
• Metadata is not private. Times, sizes and addresses exist for any hosted service. Only the content does not.

Self-hosting

Cloudflare Workers plus an R2 bucket. No database, no KV.

git clone https://github.com/raullenchai/vnsh.gitcd vnsh/worker && npm installwrangler r2 bucket create vnsh-storewrangler deploy

wrangler.toml binds R2, two native rate limiters and an Analytics Engine dataset. Earlier versions used a KV namespace for metadata and counters; it was removed after the free-tier write cap took the whole site down, and nothing needs it now — if you are following an older guide that tells you to create one, you do not. See docs/self-hosting.md.

Repository

npm test                      # every packagecd worker && npm run dev      # local worker

Four packages carry their own copy of the key schedule, because they ship independently. Each is pinned to the same test vectors: a link made by one that will not open in another reads as corruption rather than version skew, so the failure would be silent. If you change a derivation, those tests are what stops that happening quietly.

Contributing

Issues and pull requests welcome, including the ones that tell us we are wrong — the most useful contribution so far was someone reimplementing the protocol from llms.txt and reporting everything the document had failed to say.

Start with CONTRIBUTING.md. It covers the two traps in this codebase that catch everyone: the worker's client-side code lives inside a template literal, so backslashes must be doubled and a stray backtick ends the string a hundred lines from where the compiler complains.

Found a security issue? Please do not open a public issue — SECURITY.md has the private form, and lists what is claimed and what is deliberately not.

License

MIT — see LICENSE.

_{vnsh.dev · the protocol · docs · why it is built this way}