Open Source / News

God's Eye View meets Claude: we asked its MCP server 11 questions

God's Eye View 0.2 lets Claude, Codex and ChatGPT answer questions from its live data. It does this with a new MCP server. We asked it eleven questions and checked the answers at the source.

Screenshot of God's Eye View showing hurricane Rachel off Mexico, with its forecast path and cone, and hurricane Nolo at the left edge. Each has a label that Claude added.
The view Claude made on 4 October 2026, opened with "Open in God's Eye View". Claude added the Dutch labels. Screenshot: Marvin Smit. Cyclone data: NOAA National Hurricane Center / Central Pacific Hurricane Center. Imagery: Esri, Vantor, Earthstar Geographics, and the GIS User Community.

God's Eye View is a free, open-source 3D globe that shows live public data. You can now use it from inside an AI helper. Version 0.2 came out on 2 October 2026. It adds an MCP server: a small program that lets an AI helper ask God's Eye View for data. So you can ask Claude, Codex or ChatGPT about earthquakes, rockets or hurricanes. The AI then answers from the app's data. In some apps it is also meant to show the live globe.

We installed a clean copy, asked it eleven questions, and checked the answers against the original sources. The data matched. But the places it looked at did not always match the question. And in our Claude Desktop test, the globe itself did not load in the chat.

This is the fifth article in our series on God's Eye View. Earlier we wrote about what it does and what it costs, what you may share in screenshots and video and what is live, delayed and simulated.

What is new in version 0.2?

Diagram in four steps: your question, the AI helper, the MCP server on your computer, and God's Eye View fetching the public data
Diagram by Not an AI App: how a question reaches the globe. Based on the project's own setup and tools guides for version 0.2.1.

The project's list of changes calls version 0.2 "God's Eye View through Your Agent". The idea: you no longer have to open the globe and click through layers yourself. You ask an AI helper, and the AI helper asks God's Eye View.

This works through MCP, the Model Context Protocol. MCP is an open standard, a shared set of rules for connecting AI helpers to other tools. The project includes a small MCP server that runs next to the app on your own computer. Its setup guide says Claude Desktop, Codex and the ChatGPT desktop app can then show the live globe inside the chat. Claude Code and the Codex command-line tool answer with data and a link. In this Claude Code session we saw exactly that: data and a link, no globe.

The tools cover most of what the app shows:

  • earthquakes, active fires and rocket launches
  • planes and ships in an area
  • satellites overhead, and the next time a satellite passes over
  • public cameras, undersea cables and datacenters
  • weather maps, tropical storms and the height of the land
  • a combined summary of what is going on

One more tool, show_in_gods_eye_view, puts a chosen view on the globe.

Version 0.2.1 came out on the same day, with security fixes. The list of changes says the app now refuses requests from other websites to parts of the app that can cost money. It also limits the paid OpenAI and Google services to 30 and 120 requests a minute by default. We did not test those protections.

Bilawal Sidhu started the project. He published a video on 29 September with 4D rebuilds of moving animals, events and places. Those are his own experiments. We did not find them in the list of changes or the README of the public version we tested.

How we set it up

Screenshot of Claude Desktop asking: Claude wants to use Tropical cyclones from gods-eye-view, with the buttons Deny, Always allow and Allow once
Claude Desktop asked for permission before it used the tool. Screenshot by Marvin Smit, 4 October 2026.

We left our earlier test copy alone. We made a clean copy of version 0.2.1 in its own folder, without any API keys (access codes for paid data services). We followed the project's MCP setup guide:

  • Install the extra code packages the app needs. npm ci took 8.4 seconds on our laptop. The packages were already saved on it from before.
  • Build the small panel that shows the globe inside a chat: npm run build:panel, 5.1 seconds.
  • Start the app with npm run dev. Watch the port, the number at the end of the local web address. The setup guide uses 5173 in all its examples. But the app starts on 4173 by default. The MCP server also reads from 4173 if you give it no address. The port in the settings below must match the one the app runs on. We started the app on 5173 to follow the guide.
  • Tell the AI helper where the server is. For Claude Desktop, that is one block in its settings file:
{
  "mcpServers": {
    "gods-eye-view": {
      "command": "node",
      "args": ["C:\\path\\to\\gods-eye-view\\server\\mcp\\stdio.js", "--api-base", "http://localhost:5173"]
    }
  }
}

The guide asks for full file paths, because the AI helper does not load your terminal settings. On our computer, Node.js is not installed for the whole system. So we used the full path to node.exe instead of "node". We made a backup of the settings file first.

After a full restart of Claude Desktop, the tools showed up. The first time Claude wanted to use one, it asked: "Claude wants to use Tropical cyclones from gods-eye-view". The buttons were Deny, Always allow and Allow once. We chose Allow once.

One small snag: the project's setup check said "npm was not found". That was our own computer. npm was not on the list of places Windows looks for programs. After we fixed that, the check said "Ready". It did not stop anything.

Were the answers right?

Bar chart of eleven questions to the MCP server, with answer times between 24 milliseconds and 3.8 seconds. Two bars are red, where a key was missing.
Chart of our own test, 4 October 2026: the time until each answer. The two red bars are questions that need a key we had not set up.

Before we tried Claude Desktop, we wrote a small test program. It talks to the server in the same way the AI helpers do. It said hello, asked for the list of tools and then asked eleven questions. The server called itself "gods-eye-view 0.2.1" and listed 30 tools.

QuestionAnswer from the serverChecked against
Earthquakes of M4+ near Japan, last 24 hours2; strongest M5.5, Volcano Islands, 11:12:48 UTCUSGS list: same quake, same second
Active tropical stormsNolo, 100 kt, 963 hPa; Rachel, 90 kt, 965 hPaNOAA current storms: the same, from the same 15:00 UTC update
Rocket launches in the last 30 days26; latest a Falcon HeavyNot checked
Next ISS pass over Amsterdam5 October, 09:05 UTC, highest at 15°Not checked
Military planes within 100 km of San Diego1Not checked; changes by the minute
Datacenters on the map in Amsterdam81Not checked; this is what OpenStreetMap volunteers mapped, not an official count
Active fires in CaliforniaRefused: needs a NASA FIRMS keyCorrect; we had no key
Ships near the Port of RotterdamRefused: AISStream key not setCorrect; we had no key

In the table, kt (knots) is wind speed and hPa is air pressure.

Where we could check, the data matched exactly. When a key was missing, the server said so in plain words. It did not just give back an empty list. Answers took from 24 milliseconds for the datacenter list to 3.8 seconds for the launch list.

On purpose, we left out every question that uses the general flight data from the OpenSky Network. We explain the licence reasons in our third article. The military planes question uses adsb.lol instead.

Where it goes wrong: the area behind the answer

Screenshot of God's Eye View looking down on Europe and Africa, while its storm panel lists hurricanes Rachel and Nolo, which are in the Pacific
The storm tool's default view: the panel lists Rachel and Nolo, but the camera looks at Africa. Our screenshot, 4 October 2026, no keys. Cyclone data: NOAA National Hurricane Center / Central Pacific Hurricane Center. Imagery: Esri, Vantor, Earthstar Geographics, and the GIS User Community.

Two answers were correct, but they looked at a different place than a reader would expect.

"Submarine cables in the Netherlands": 178 undersea cables. The list included 2Africa and the Africa Coast to Europe cable. So we checked how the app understood the place name. The answer was clear. "Netherlands" became the whole Kingdom of the Netherlands, including its Caribbean islands. The area ran from 53.7° north down to 11.8° north, and from 7.2° east to 70.3° west. That covers a large part of the Atlantic Ocean. The count is right for that box. But it is not what most people mean by the Netherlands.

The storm view pointed at Africa. Every answer comes with a suggested view for the globe. For the storms, that view looked straight down at 20° north and 0° east, over West Africa. But both hurricanes were in the Pacific Ocean. You can see the result in the image above. The panel lists Rachel and Nolo, but the camera shows Africa. If an AI helper passes that view on unchanged, you look at the wrong side of the planet.

Neither is a data error. Both remind us that an AI tool's answer depends on how it understood your question. Ask for an area that is as exact as you can make it, or check the area the tool reports back. The server's own instructions to AI helpers say something similar. Results have a limit, so check the total before deciding there is nothing more.

Did the globe show up in the chat?

Screenshot of God's Eye View zoomed in on hurricane Rachel, with its forecast path marked at 12 to 72 hours and a shaded cone showing where it may go
The same data with an area added: hurricane Rachel with its forecast path and cone. Our screenshot, 4 October 2026, no API keys. Cyclone data: NOAA National Hurricane Center / Central Pacific Hurricane Center. Imagery: Esri, Vantor, Earthstar Geographics, and the GIS User Community.

This is the main new feature of version 0.2, and in our test it did not work. We asked Claude Desktop: "Show the active tropical cyclones in God's Eye View." Claude asked permission for the storm tool. Then it used show_in_gods_eye_view. Where the globe should have been, the panel said: "God's Eye View could not load here: Only the God's Eye View panel may make this request."

The rest of the answer was useful. Claude wrote a short summary with a table of both hurricanes: their position, wind, air pressure and movement. It also noted that this source only covers the Atlantic and the eastern and central North Pacific. All values matched NOAA's update of 15:00 UTC.

The panel also had a button: "Open in God's Eye View". It opened the view in the app, framed on both hurricanes. Claude had added a label to each one. That is the image at the top of this article. The labels were in Dutch, "Orkaan Nolo" and "Orkaan Rachel", even though the question was in English. Claude chose that itself.

Why did the globe not load? We looked in the code. The error comes from a check in the server. Each copy of the MCP server makes its own secret key when it starts. The globe panel must send that key with its requests. When we looked afterwards, Claude Desktop had started two copies of the server, in the same second. Suppose the panel came from one copy and its requests went to the other. Then the keys would not match, and this exact message would appear. That is our best explanation, but we have not proven it. It may also depend on the mode we used, Cowork, or on our Windows setup.

When we added an area to the request, the same server did zoom in on hurricane Rachel in the app. It showed the forecast path from 12 to 72 hours, and the shaded cone that shows where the centre may go. That is the picture in this section. The cone and the path are NOAA's forecast as the app received it. The app did not work them out itself.

Questions people ask

List of questions people ask Google about this topic, with the 3 questions we answer marked
"People also ask" questions from Google (US), 9 October 2026, via DataForSEO. Chart by Not an AI App.

What is the Model Context Protocol (MCP) and how is it used in AI applications?

MCP is an open standard: a shared set of rules for connecting AI helpers to other tools. A small program, the MCP server, lets an AI helper ask a tool for data. With God's Eye View 0.2, you can ask Claude, Codex or ChatGPT about earthquakes, rockets or storms, and the AI answers from the app's data.

How to add MCP servers to Claude?

For Claude Desktop, you add one block to its settings file. It names the server, the command that starts it and full file paths, because Claude does not load your terminal settings. Then restart Claude Desktop fully. In our test on 4 October 2026, the tools showed up, and Claude asked permission the first time it used one. See the MCP setup guide.

When to use MCP?

When you want an AI helper to look things up in another tool for you. With God's Eye View, you no longer have to open the globe and click through layers yourself. But check the answer: in our test, two answers looked at a different area than the question meant.

How we checked

Diagram of our test in three steps: install a clean copy without keys, ask eleven questions through MCP, check the answers against USGS and NOAA
Diagram by Not an AI App: how we tested the MCP server.

We read the project's list of changes (CHANGELOG), README, MCP setup guide and tools guide at version 0.2.1 (code version aa16b7c). Settings and tools can change with any update.

We installed that version on 4 October 2026 in a separate folder. We used the Windows laptop described in our first article, without any API keys set up. Our test program talked to the server using MCP, through the server's standard text input and output. It asked the eleven questions in the table. We compared the earthquake with the USGS list and the storms with NOAA's list of current storms. We fetched both on the same evening. We checked how the app understood "Netherlands" by using its own place search.

The owner of this site, Marvin Smit, did the Claude Desktop test. He used Claude Desktop with Opus 5.5, in its Cowork mode. The screenshots from Claude Desktop are his. We only made them smaller. Afterwards we counted how many copies of the MCP server Claude Desktop had started. We took the other screenshots ourselves, from the links the tools gave back. We only made them smaller. We made the charts and diagrams from our own test file. None of the images is AI-generated.

Our review covers the MCP server with Claude Desktop and our own test program. It does not cover ChatGPT, Codex, the web (HTTP) version of the server, features that need a key, or the security fixes in version 0.2.1. This is an editorial review of the sources we read and the tests we ran.