Works today in ChatGPT desktop & Chrome 149

The website builder built for humans and agents.

Build visually. Direct your agent. Publish an agent-ready website. A real visual editor where a person and an AI agent author the same live project through WebMCP — structured context, reviewable proposals, and a human approval gate before anything ships.

Live demo storefront in your browser. Nothing to install, no account.

The WEAVE editor: a ceramics storefront on the canvas, with the Agent panel showing a live WebMCP connection, an agent readiness score of 100, the available tools and a feed of agent tool calls.
document.modelContext — current W3C API 48 typed tools, real annotations 10,093 tests passing AGPL-3.0 — your code, no lock-in

How it works

One project. Two authors.

The agent never guesses from pixels and never edits behind your back. It reads the same structured model the canvas renders from, and proposes work you approve.

Human

Edit on a real canvas

Drag, type, restyle. Every edit becomes a new project revision.

Agent

Read structured context

Typed sections, stable element ids, your selection — not a DOM dump.

Agent

Propose a change set

Related edits arrive as one reviewable proposal. The canvas doesn’t move.

Human

Amend, skip, apply

Rewrite a value, drop an operation, apply the rest. One undo reverses it all.

WEAVE

Gate the publish

An agent can request a publish. Only your click performs one.

Change sets

Transactions, not one-off edits.

“Make this homepage feel more premium” becomes one proposal containing several operations — not five edits fired at your canvas one at a time.

  • See every operation’s before and after
  • Rewrite any value the agent chose
  • Skip the parts you don’t want
  • Accepted work commits atomically as one revision
  • A single undo reverses the whole transaction
The proposal review panel: an agent proposal titled Make the homepage feel more premium, showing three operations with their current and proposed values, one marked edited by you, and an Apply 3 changes button.

Safety

The agent can ask. Only you can ship.

There is no code path from a tool call to a deploy. A publish request opens an approval card showing the revision and exactly what changed since the last one.

  • Destructive tools are annotated as destructive
  • Proposals pinned to a revision — stale ones are refused
  • Three validation layers before anything touches your project
  • No eval, no arbitrary code, no javascript: URLs
The Agent panel showing an Agent requested publish card with the revision, a summary of what changed, and Approve and publish alongside Cancel buttons.

Inspectable

Every claim is checkable, live.

The Inspector reports what is actually registered right now — read from the runtime, not from our documentation.

  • Which global carries the runtime, and which API members exist
  • Every tool’s JSON schema and annotations
  • Which tools are exposed vs hidden, and why
  • The last invocation each tool received
The WebMCP Inspector dialog listing the detected runtime host, capability pills for registerTool, unregisterTool, getTools and more, and every tool with its access level and exposure state.

Efficiency

Structure costs less than pixels.

Measured on this build’s demo page. An agent reading structured context instead of the raw DOM spends roughly a tenth of the tokens — and every id it gets back is directly actionable, so there is no second “find the element” step.

~3.1ktokens to read the whole page12,335 chars structured
~30ktokens for the same page as raw DOM119,875 chars
10×less to read, every timemeasured, not estimated
1proposal replaces five verify loopsone commit, one undo

The surface

Nine typed tools.

Registered with the browser’s model-context runtime. Descriptions state product intent; annotations declare real consequences. Element tools appear only while you have something selected.

ToolAccessWhat it does
weave_get_contextreadPage, revision, selection, typed sections, bounded element tree
weave_get_selectionreadFull detail for the element you have selected
weave_add_sectionwriteInsert a designed section from the shared library
weave_update_elementwriteText, name, visibility, safe styles and attributes
weave_move_elementwriteReorder or re-parent a section or element
weave_delete_elementdestructiveRemove an element and its children
weave_propose_changeshuman reviewSubmit several edits as one reviewable proposal
weave_validate_sitereadReal findings plus an explainable agent-readiness score
weave_publish_sitehuman approvalRequest a publish — only your click performs one

Where this actually stands

  • WebMCP is an emerging standard. If your browser exposes no runtime, the editor says so plainly — the tools stay usable through the built-in developer console, which calls the identical implementations.
  • Agent context is pull-based: WEAVE bumps a revision on every change and agents re-read. It does not pretend to push page content the shipping API cannot push.
  • “Publish” means persist and download a complete, agent-ready bundle. No third-party deployment provider is contacted or simulated.
  • The runtime shipped with a published site is read-only plus same-site navigation, by design.

Open it and give your agent something to read.

A live storefront, nine tools, and an approval gate. No install, no account.