Connect an AI agent to Alterno Spatial Review

Requirements and compatibility

Live tools require a browser and agent host that expose the page’s document.modelContext WebMCP API. Ordinary browsing or support for a remote MCP server does not establish WebMCP support. This site does not advertise a remote HTTP MCP endpoint.

This guide does not claim that every version of any named browser or coding agent supports the integration. Check for page-defined review.* tools in the actual host. If those tools are absent, use the editor’s manual controls and feedback JSON export.

Open a Spatial Review-enabled website in the editor, then use Connect AI and copy the operational prompt for your work mode. Editor-first mode awaits new review activity; agent-first mode works on demand. The host must actually keep the awaited activity call running to listen continuously.

Ten native tools

• review.connect — join a review and obtain readiness and resume information.
• review.disconnect — end presence and cancel waits without deleting saved reviews.
• review.report_work — show working or idle presence; silence expires after 30 seconds.
• review.read_activity — read events, page history, or await new activity.
• review.read_feedback — read requests and discussion references with bounded pagination.
• review.post_message — post to shared chat or reply to a specific discussion.
• review.set_request_status — address or reopen a request using its revision.
• review.discover_tools — retrieve complete typed capability contracts.
• review.call_read_tool — invoke a discovered read contract with its tool token.
• review.call_write_tool — invoke a discovered write contract with its request ID.

A minimal workflow

• Connect before reading heavy context. Read activity and feedback to understand the user’s request.
• For ordinary chat, feedback, and status actions, use native tools directly. Discover only additional capabilities needed for the task: inspect, scene, assets, experience, representation, discussions, views, or export.
• Follow every discovery cursor. Reassemble ordered UTF-8 contract chunks and verify their byte counts and SHA-256 before using the delivered token. Cache a complete contract within its origin and installation scope.
• Use get_context for readiness and the human view. Its editor rootRef identifies an editor, not an inspectable target. Use find_targets to obtain object, asset, or sequence target references before read_targets; an editor root passed to read_targets returns REF_KIND_MISMATCH.
• Read the actual target references and edit base before a spatial write. Keep newer user edits and refresh affected state after a revision conflict.
• Use a stable requestId for identical retries after uncertain results. A changed payload needs a new ID.
• Reply in the relevant discussion. Mark a request addressed only when its requested work is handled; a status change does not prove that website source was implemented.
• When the user requests source changes, apply them through your external authoring tools. Explicit navigation is separate from inspecting or editing hidden workspaces.

Limits and fallback

Scene, Assets, and Experience share one review connection, while independent saved reviews retain separate documents and conversations. Do not guess opaque references or infer source ownership from an asset name.

Isolated capture has renderer and freshness limitations. Report unavailable or stale captures explicitly. A review image is not proof of pixel-identical website lighting or post-processing.

If WebMCP is unavailable, export feedback JSON. It covers all three editor scopes, including saved or unvisited scopes, without requiring a live agent connection.