Your agent built it. Open it anywhere.

Tell the agent. Get the link.

Press play to watch a stand-in agent run the prompt, or copy it and paste it into your own. It uses the docs as command reference, keeps the page high-level, and hands back a no-login link spelling out how you should be using agent-paste. When you want to keep it, revise it, or run JavaScript, ask the agent to claim it and sign in free.

When you’d reach for it.

1. 01

   Your agent wrote a research brief or analysis you need to send up.

   Hand your boss a link, not a 4,000-line chat scroll. It opens in any browser, formatted, no tool of theirs required.

2. 02

   You built something in one agent and the next person lives in another.

   Publish from Cursor, and a teammate on ChatGPT or Claude reads the work instead of re-deriving it. The handoff carries the files, not a copy-paste.

3. 03

   The agent keeps revising and people already have the link.

   Everyone you sent it to sees the newest version on their own, no reload and no re-send. One stable link for the whole back-and-forth.

4. 04

   Your agent generated a page that is meant to actually run.

   A self-contained dashboard or interactive report runs from a safe, isolated origin, so you can open what the agent wrote without standing up a server.

If the agent has a shell, it runs the publish itself. With no account it returns an unlisted URL to a static page in seconds.

To keep the work, run JavaScript, and own it, log in for free once over browser OAuth, then publish.

Browser OAuth provisions a scoped key on your machine. No key to copy or paste.

In a web chat with no terminal, like ChatGPT, Claude, or Gemini, add the server once and sign in free. The agent then publishes, reads, and shares from there. The accountless path above is the shell route; chat connects after a free login.

Add it as a remote MCP server in your client. See the setup docs.

1. 01

   One link, opens anywhere

   Hand it to a person or pass it to another agent. People get a clean View in the browser; agents get an Agent View, structured JSON with the file tree, metadata, and signed per-file URLs, so the next agent reads the work instead of scraping it. One stable Artifact, the same across CLI, MCP, and the dashboard.

2. 02

   Move work between tools that don't talk

   Work made inside one tool stays walled in: vendor surfaces are auth-locked with no machine-readable way out. agent-paste is the neutral layer between them. An agent in Cursor publishes; a teammate in ChatGPT, or you in a browser, picks it up.

3. 03

   The link stays current, until you cut it

   Leave the link open and every viewer advances to the newest Revision on its own, no reload, no re-send. Share through a revocable Access Link and pull it back any time without deleting the work. Artifacts expire under your Workspace Auto Deletion policy: a handoff, not a vault.

4. 04

   Safe to open what your agent wrote

   Generated pages are untrusted by construction, so they run from an isolated Content Origin: private storage, short-lived signed tokens, platform-derived MIME types, a strict execution policy, and per-artifact lockdown. You can host what an agent generated without it touching your account.

Try it with no account. An --ephemeral publish takes text, images, and static HTML with no JS and returns a link you can open right away. When you want to keep it longer or run JavaScript, claim it: a free login that moves the work into your account.