marklayer-mcp
MCP (Model Context Protocol) server that bridges MarkLayer annotations to AI coding agents.
When users annotate any webpage with MarkLayer, your agent receives the comments as a structured work queue: it can acknowledge, resolve, dismiss, and reply to each one — and the human sees the status updates live. It also works the other way: point an agent at a room and ask it to review the page, and it can leave its own comments and copy-edit suggestions for the human to triage.
Install
A standard stdio MCP server — no Claude-specific wiring, works with any MCP client (Claude Code, Cursor, Windsurf, VS Code, Codex CLI, etc.).
claude mcp add marklayer -- npx -y marklayer-mcp
claude mcp add marklayer -- npx -y marklayer-mcp --room https://marklayer.app/s/abc123
For any other MCP client, add a stdio server entry pointing at the same command (adjust the config file/key to your client — Cursor's ~/.cursor/mcp.json, Windsurf's mcp_config.json, VS Code's .vscode/mcp.json, etc.):
{
"mcpServers": {
"marklayer": {
"command": "npx",
"args": ["-y", "marklayer-mcp"],
"env": { "MARKLAYER_AGENT": "cursor" }
}
}
}
Set MARKLAYER_AGENT (or --agent) to identify which agent this is — it's what humans see as the comment author and the "assigned to" badge, so a non-Claude client should not leave it at the default.
Usage
In your agent prompt:
Watch my MarkLayer annotations. For each one, acknowledge it, make the fix, then resolve it with a summary.
The agent will call marklayer_watch_annotations in a loop and process incoming feedback as it arrives.
Tools
marklayer_connect_room | Connect to a room by share URL or bare id. |
marklayer_room_info | Page URL, viewport width, timestamps. |
marklayer_list_annotations | List annotations, optionally filtered by status. |
marklayer_get_annotation | Full detail + reply thread for one annotation. |
marklayer_watch_annotations | Block until new annotations arrive, return a batch. |
marklayer_acknowledge | Mark in-progress and tag with this agent. |
marklayer_resolve | Mark resolved, optionally posting a reply with the summary. |
marklayer_dismiss | Mark dismissed with a reason the human will see. |
marklayer_reply | Post a reply without changing status (e.g. clarifying questions). |
marklayer_create_annotation | Leave a new comment yourself — proactive feedback rather than a response. |
marklayer_suggest_edit | Propose an exact copy/grammar fix for a piece of text, as a diff the human can accept. |
Options
--room <url-or-id> Connect at startup; otherwise call marklayer_connect_room.
--api-base <url> Override worker URL (default https://marklayer.app).
--agent <name> Identifier shown to humans (default "claude-code").
Equivalent env vars: MARKLAYER_ROOM, MARKLAYER_API_BASE, MARKLAYER_AGENT.
Troubleshooting
"Connection closed" on first start. npx -y downloads the package on the
first run, and a cold download can outlast the MCP client's startup timeout.
Prime the cache once (npx -y marklayer-mcp --help) and reconnect, or install
it up front with npm i -g marklayer-mcp and point the client at the binary.
Tools return "room is not connected". The WebSocket dropped — rooms are held
open by a Durable Object and a long-idle agent can be disconnected. Call
marklayer_connect_room with the same URL to reattach.
Creating rooms programmatically
This server consumes an existing share link. To mint one (or many) from code — e.g. seed a room per page in a batch of URLs before pointing the agent at it — POST directly to the public HTTP API:
curl -X POST https://marklayer.app/api/$ID \
-H 'Content-Type: application/json' \
-d '{"ops":[],"url":"https://example.com/page-1","width":1440,"expires_in":2592000}'
$ID is caller-supplied (use nanoid / crypto.randomUUID() — it's the access token). No auth, no SDK. Full details: https://marklayer.app/llms-full.txt