Skyelight MCP
The Skyelight MCP server connects a coding agent to your Skyelight projects. The agent can list open threads, read a thread with its page, element, screenshot and source location, and reply on the thread when it has made a change. It can also publish a prototype or a deployed app to a project, so people can review it.
MCP (Model Context Protocol) is an open standard for connecting agents to tools. Any MCP client can connect to Skyelight, including Claude Code, Cursor, Codex, Windsurf, claude.ai and ChatGPT.
What an agent can do
The server provides 18 tools. They cover six kinds of work:
- List the open threads on a project, filtered by page, type, status or assignee, or the threads assigned to you.
- Read one thread in full: the replies, the page URL, the anchor, the screenshot and attached images, and the source file and line when the site uses Skyelight Build.
- Reply on a thread with what changed and a link to the pull request, then resolve or defer it.
- Check whether something has already been reported, and merge duplicate threads.
- Summarize a project: what changed since a given time, the feedback on one file, and what the project resolved and decided.
- Publish something for review: the code of one page or component, which
Skyelight hosts on
skyelight.page, or the address of an app that is already deployed. See Hosted prototypes.
Each tool’s arguments and limits are in the tools reference.
Connection options
| Remote server | Local server | |
|---|---|---|
| How it runs | Your agent connects to https://app.skyelight.ai/mcp over HTTP | The @skyelight/mcp package runs on your machine |
| Sign-in | OAuth, once, in your browser | A personal token (sky_…) in your agent’s config |
| Stored on disk | Nothing | The token |
| Use it when | Your agent supports OAuth, which most do, including claude.ai and ChatGPT | You can’t sign in through a browser |
Both options provide the same tools. Use the remote server unless you have a reason not to. Connect your agent covers both.
In Claude Code, the Skyelight plugin sets up the remote server for you. Skyelight is also listed in the official MCP Registry, with both options.
Requirements
To use agent tools, you need:
- An owner, admin or collaborator role in the workspace. Reviewers can’t use agent tools on any plan.
Every plan gets every tool, Free included. Pins and threads are unlimited on every plan. On Free, your agent gets 25 agent threads a month:
- Looking counts one per conversation. The first time a conversation calls Skyelight, with any tool, listing included, it counts one. Each new session in Claude Code, Cursor or claude.ai is a new conversation. ChatGPT, the npm package and the REST API count one a person a day.
- Each thread your agent works counts one more, the first time in the
month
get_item,post_update,set_status,assignormerge_itemstouches it. After that the agent can keep working it as much as it needs. - So asking “what feedback do I have?” and then fixing one pin is 2.
- Nothing gives one back: resolving or deleting a thread, or ending the conversation.
- At 25, conversations and threads already counted keep working, and anything new is refused with the date the count starts again. Pro and above have no limit.
On Free, list_workspaces says how many threads are used and left this month.
Free hosts one prototype, with any number of versions.
If a requirement is missing, the agent still lists the workspace, marked with what the agent can’t do there, and its tools return an error that names the missing requirement. See Plans and billing and Roles and permissions.
Agents act as you
Skyelight has no service accounts. The agent uses your sign-in or your token, so it has the same access as your account in each workspace.
- Replies the agent posts appear under your name, with a mark that names the client, such as Claude Code or Cursor.
- Mentions, assignments and status changes send the same notifications as they do in the app.
Thread text comes from the people reviewing your app, and it can contain instructions. Skyelight marks every tool as read-only or write, and most clients ask before running a write tool. Read what the agent proposes before you approve it.