For coding agents
The site's numbers are a dataset first: 22 devices and 135 readings, version 1.0.0 of 2026-11-17. Connect an agent to the MCP server, install the agent skill, fetch the files below, or fetch a Markdown copy of any page.
MCP server
https://screens.guide/mcp is a remote MCP server over Streamable HTTP, with no sign-in. Add it to Claude Code with claude mcp add --transport http screens-guide https://screens.guide/mcp, or give the URL to any MCP client. Its tools read the dataset and change nothing:
| Tool | What it does |
|---|---|
list_devices | List the measured devices |
get_device | Get a device's readings |
compare_devices | Compare devices |
check_layout | Check a layout against a device |
convert_units | Convert between pt, dp and px |
- Lengths are in the device's unit and positions in window coordinates, as in the files.
check_layouttakes frames in the same coordinates and reports each one that reaches outside the window, into an inset, past a rounded corner, under a system bar, across a fold or into a reserved region, with the numbers. - Each client address may send 100 requests a minute and 1,000 a UTC day. A request over either limit gets HTTP 429 with
Retry-After, the seconds until the limit resets. - The server keeps no session and answers each request with one JSON response; it opens no event stream.
Agent skill
The screens-guide skill tells an agent where to read the numbers, which layout rules the measurements support, with their numbers, and how to review a screen against a device. It carries a snapshot of each device's current readings, references/devices.md, for an agent without network access. Read it at /skills/screens-guide/SKILL.md, or download the folder as screens-guide.zip.
Install
The Claude Code plugin installs the skill and the MCP server together. Elsewhere they install apart: the skill from the repository or the zip, the server from its URL.
Claude Code
- The plugin, the skill and the MCP server:
/plugin marketplace add modaal-agent/screens-guide, then/plugin install screens-guide@screens-guide. On Claude Code 2.1.275 or later one line does both:/plugin install screens-guide --marketplace modaal-agent/screens-guide. - Update with
claude plugin marketplace update screens-guide, thenclaude plugin update screens-guide@screens-guide;/reload-pluginsapplies it to a running session. Claude Code installs an update only when the plugin's version changed. - The MCP server alone:
claude mcp add --transport http screens-guide https://screens.guide/mcp. - The skill alone: put the zip's
screens-guidefolder in~/.claude/skills/for every project, or in a project's.claude/skills/. Claude Code picks it up without a restart.
claude.ai
- The skill: download screens-guide.zip, open Customize, then Skills, click +, then Create skill, then Upload a skill, and choose the zip. Skills need code execution, which Settings turns on under Capabilities.
- The MCP server: Customize, then Connectors, Add, Add custom connector, with the URL
https://screens.guide/mcp. The server needs no authentication.
The Claude API
Upload the skill's files to POST https://api.anthropic.com/v1/skills, each as a files[] field named screens-guide/<path>, then name the returned skill_id in a Messages request's container.skills, beside the code execution tool. The steps are in Anthropic's skills guide.
Other agents
Codex, Cursor, GitHub Copilot and Gemini CLI read the same skill format. Each step below follows the agent's documentation as read on 2026-11-18.
- Any of them, with the cross-agent `skills` CLI:
npx skills add modaal-agent/screens-guide --skill screens-guideinstalls the skill into each agent it finds on the machine;-a <agent>names one, and-ginstalls it for the user instead of the project.npx skills add https://screens.guide/skills/screens-guide.zipinstalls it from the zip. (docs) - Codex: skills load from
.agents/skills/in the project and~/.agents/skills/. The server:codex mcp add screens-guide --url https://screens.guide/mcp. (skills, MCP) - Cursor: skills load from
.agents/skills/or.cursor/skills/in the project, and~/.cursor/skills/. The server:{"mcpServers": {"screens-guide": {"url": "https://screens.guide/mcp"}}}in.cursor/mcp.jsonor~/.cursor/mcp.json. (skills, MCP) - GitHub Copilot:
gh skill install modaal-agent/screens-guide screens-guide, with GitHub CLI 2.90.0 or later; skills load from.github/skills/or.agents/skills/in the project, and~/.copilot/skills/. The server in Copilot CLI:copilot mcp add --transport http screens-guide https://screens.guide/mcp. (skills, MCP) - Gemini CLI:
gemini skills install https://github.com/modaal-agent/screens-guide.git --path skills/screens-guide. The server:gemini mcp add --transport http screens-guide https://screens.guide/mcp. (skills, MCP)
Files
| URL | What |
|---|---|
/data/v1/devices.json | the whole dataset: every device and every reading |
/data/v1/devices/<id>.json | one device, with the dataset's version and the runs its readings name; for example /data/v1/devices/iphone-17.json |
/data/v1/devices.csv | one row per device, screen, orientation and runtime; nested fields joined with _, lists and missing as JSON text |
/data/v1/schema.json | the JSON Schema (draft 2020-12) of devices.json; a device file's device follows its $defs/device |
/data/v1/version.json | the dataset version, its date, and the counts of devices, runtimes and readings |
/llms.txt | an index of the site for language models |
/llms-full.txt | every device's current reading as Markdown tables, and the guide |
Every page has a Markdown copy at its path plus .md: /devices/iphone-17.md for /devices/iphone-17, and /index.md for the home page. A request for the page itself with Accept: text/markdown gets the copy, and the HTML names it in a Link header. The files and the copies allow requests from any origin.
Reading the data
- One reading per device, runtime build, screen and rotation. A device's
currentnames the build to show first: the newest release build with a reading in each screen and rotation. - Lengths are in the device's
unit: points on iOS, dp on Android. Positions are window coordinates from the top left, already rotated for the reading's orientation. - A value no run measured is
null, and the reading'smissinggives the reason for that field. Never fill it from another device. - A reading on a beta build has
runtime.prereleasetrue; its values can change by the release. cornerRadiihas two sets,profileandmeasured, each with itssource. On the iPads on iOS 27.2 and 27.3 they differ.
License and citation
The dataset is licensed under Creative Commons Attribution 4.0: copy, change and redistribute it, commercially too, with credit. Credit it as "screens.guide, measured by Modaal" with a link to https://screens.guide, and name the version you used, 1.0.0. The code and pages are MIT-licensed.