Assistant

Versión en español

Assistant for managers

The assistant is a simpler way to use the MCP server: a chat in its own window, on your computer, with the language model of MonitorEO, Claude Sonnet 5.5 by Anthropic, which runs in the cloud through MonitorEO, with nothing to install and no account or key; or, if you prefer, with an open model that runs on your own computer (with Ollama), which answers worse (see Which model). You ask in plain language for what you want to compute, for example "Monthly NDVI in Doñana National Park in 2024". The assistant looks for the area, chooses the variable and the sensor, starts the computation, shows the Earth Engine maps (click on a pixel to see its value and its time series) and gathers the results into a report to print or save as PDF.

You can write in English or in Spanish: the assistant answers in the language of each question (a Spanish question gets a Spanish answer). The examples in this page are in English; they work the same in Spanish.

The assistant does not interpret the results. The figures, charts and maps go straight to your screen and to the report. The model only receives an acknowledgement ("computed and shown"), without the values, so it cannot describe them or draw conclusions. If you ask, it answers that interpretation is a job for technical staff. It does advise on which variable, sensor, period or tool to use, with the information of the MonitorEO catalog, answers questions about MonitorEO with this manual (the wiki) and, if something cannot be done with the assistant, tells you which other MonitorEO tool does it.

It uses the same tools as the MCP server, with the same results. The difference is who it is meant for:

	Assistant	MCP with Claude or another assistant
Language model	That of MonitorEO in the cloud (Claude Sonnet 5.5) or an open one on your computer (Ollama)	The one of your client: Claude, etc.
The results	Shown, not interpreted	The assistant reads and explains them
Reports	HTML to print or save as PDF	Whatever you ask the assistant for
Tools	The computation ones, the GeoTIFF downloads and the exports to Google Drive	All of them, also publishing in the viewer (administrators)
Installation	An installer (double click on Windows, one command on macOS and on Ubuntu); with a model on your computer, the assistant downloads it the first time	Python package and an MCP client

Where it computes

• The computations run in Google Earth Engine, with the shared account of MonitorEO or, if you type your project in Settings, with your Google account and your Google Cloud project (they count against its quota), as in QGIS, Python or the MCP. Note: under Google's terms, the operational management of a public administration needs a commercial Earth Engine project, not a non-commercial one like that of the shared account (see Getting started).
• The language model, by default that of MonitorEO: Claude Sonnet 5.5, by Anthropic, through the server of MonitorEO (no account or key needed). It only receives the text of the conversation, which goes to Anthropic, which does not use it to train its models (its terms). If the credits of MonitorEO at Anthropic run out, Gemini 3.8 Flash, by Google, answers; with the paid tier of its API, it does not use the conversation to improve its products either (its terms). With a model on your computer (see Which model), the conversation does not leave it.
• Your study areas are sent to Earth Engine to compute on them; the model only sees their name.

The window reminds you in its header, with the project and the model it uses. Below, a bar tells you which part of your Earth Engine computing EECUs of this month you have spent (with the shared account, your research EECUs: 25 EECU-h a month, for research only) and Check asks the question again; if you have less than 10 % left, the assistant says so in the conversation, and when they run out you have to switch to your own account. With your own project, choose its level in Settings (see Concepts).

Installation

The installer does everything, with no administrator rights and no need to know how to use a terminal.

1. Earth Engine account and project, only if you want to compute with your own: see Getting started. Without them, the assistant uses the shared account of MonitorEO.

2. Open the installer:

   • macOS: open the Terminal (press Cmd+Space, type Terminal and Return), paste this command and press Return:

     /bin/bash -c "$(curl -fsSL https://raw.githubusercontent.com/monitoreo-obsnev/monitoreo/main/installer/macos/instalar.sh)"

     It is the only time the Terminal is needed. macOS blocks unsigned programs downloaded with the browser, and the installer is not signed; what this command downloads is not blocked. Step by step:

     

   • Windows 10 or 11: download asistente-monitoreo-windows.zip, open it and double-click Install the MonitorEO assistant.bat. As it is not signed, Windows may warn "Windows protected your PC": click More info and Run anyway.

   • Ubuntu 22.04, 24.04 or 26.04 LTS (on a PC or on ARM): open the Terminal (press Ctrl+Alt+T), paste this command with Ctrl+Shift+V and press Enter:

     bash -c "$(wget -qO- https://raw.githubusercontent.com/monitoreo-obsnev/monitoreo/main/installer/linux/instalar.sh)"

     It does not ask for your password, because it does not touch the system. It also installs the window of the assistant (Qt WebEngine), so on Ubuntu it is about 0.6 GB to download and 2.5 GB on disk.

   A window opens that tells you what it is doing. It downloads and installs Python, Earth Engine, GDAL and MonitorEO in the MonitorEO folder of your user (about 0.2 GB to download and 1 GB on disk, and on Windows 0.3 and 1.7 GB; on a Mac, one minute). You do not have to type anything.

3. When it finishes, the assistant opens in its own window, and MonitorEO Assistant is left on the Desktop (an app on macOS, a shortcut on Windows and on Ubuntu, where it is also among the applications: press the Super key and type MonitorEO) to open it again with a double click, with no Terminal or console. The installer window can be closed. To close the assistant, close its window (or, on macOS, Cmd+Q): the engine of the models closes too. If it is computing, it asks first, because the computation would be lost.

4. In Settings, the first time:

   • Earth Engine account: if you do nothing, the assistant computes with the shared account of MonitorEO. To compute with your own (and export to Google Drive), click Use my Google account, which opens the browser to sign in as earthengine authenticate does (once per computer; it also works for QGIS, Python, R and the MCP), and type in Your project (optional) the one of your account registered in Earth Engine (the one you see at the top right of the Code Editor), then Save. To go back to the shared account, leave the project empty and click Save. With your account, Change account signs in with another one and Log out signs your Google account out of Earth Engine on this computer (it revokes its permission and deletes its credentials, also for QGIS, Python, R, the command line and earthengine): the assistant goes back to the shared account.
   • Language model: click Install. The assistant downloads the model (qwen3:8b, about 5 GB) and the engine that runs it, Ollama (0.2 GB on macOS and 1.5 GB on Windows and Linux).

   When everything is ready, Settings folds and says "· all ready". It is saved in ~/MonitorEO/assistant.json.

The first time, about 5 GB have to be downloaded (5.4 GB on macOS and 6.7 GB on Windows and Linux, with the engine), so a good connection helps. It happens only the first time: afterwards the assistant starts without downloading anything, except the MonitorEO updates, which are small.

If it does not work this way

• macOS, without the Terminal: download asistente-monitoreo-macos.zip, open it and double-click Install the MonitorEO assistant.command. macOS does not let you open it the first time ("Apple could not verify…"): click Done, go to System Settings > Privacy & Security, scroll down to the warning about the installer and click Open Anyway, which asks for the password of an administrator of the Mac.

• Windows, if the computer does not let you open the zip (SmartScreen or a policy of your organization): open the Command Prompt (Start menu, type cmd and Enter), paste this command and press Enter:

  curl -fsSLo "%TEMP%\instalar-monitoreo.bat" https://raw.githubusercontent.com/monitoreo-obsnev/monitoreo/main/installer/windows/instalar.bat && "%TEMP%\instalar-monitoreo.bat"

• Ubuntu, if the Desktop icon does not open the assistant (it shows a cross or opens a text): right-click it and choose Allow Launching, or open it from the applications (Super key and MonitorEO). If the Terminal says that it cannot find wget, install it with sudo apt install wget (it asks for your password) and paste the command again.

What it installs and where

Everything stays in your user folder, without touching the system:

Folder	What it is
MonitorEO/entorno	Python, Earth Engine, GDAL and MonitorEO (installed by the installer, with micromamba)
MonitorEO/micromamba	The program that installs the conda-forge packages (on Windows, with the DLLs of the Visual C++ runtime that it needs, if Windows does not have them)
MonitorEO/ollama	The engine of the models on your computer, if you choose one and did not have Ollama
MonitorEO/wiki	This manual, to answer questions (updated every day)
.ollama/models	The language models on your computer, if you choose one (about 5 GB)
MonitorEO	Also: the settings (assistant.json), your study areas (capas), the reports (informes) and the results

To uninstall it, delete MonitorEO/entorno, MonitorEO/micromamba, MonitorEO/ollama and .ollama, and MonitorEO Assistant (or Asistente de MonitorEO, if it comes from an earlier version; reinstalling replaces it) from the Desktop (the results, the layers and the reports are in the rest of MonitorEO). To install it again, delete MonitorEO/entorno and open the installer again.

On Windows, while it installs, the packages are in C:\MonitorEO-pkgs, which it deletes at the end: in your user folder, the path of some of their files would pass the 260 characters that Windows allows. If an installation stops halfway, you can delete that folder.

On Ubuntu, the installer also puts MonitorEO Assistant among the applications (.local/share/applications/monitoreo-assistant.desktop, with its icon, .local/share/icons/monitoreo-assistant.png), which opens the assistant with MonitorEO/entorno/bin/monitoreo-assistant: to uninstall it, delete those two files too.

With the conda environment

If you already have the MonitorEO environment (Getting started), you do not need the installer:

conda activate monitoreo
monitoreo assistant

It opens in its own window (with pywebview, which is in the environment); if the computer cannot open it, in the browser. To close it, close the window or press Ctrl+C in the terminal; in the browser, use Close the assistant, at the top right (closing the tab does not close it).

On Linux, the window needs Qt WebEngine, which is not in the environment (without it, it opens in the browser). To have it: conda install -n monitoreo -c conda-forge pyqt6-webengine qtpy, and open it with QT_API=pyqt6 QTWEBENGINE_DISABLE_SANDBOX=1 monitoreo assistant, as the Ubuntu installer does (with the sandbox of Chromium, the page may stay blank).

Which model

In Settings, Language model has two options:

• The model of MonitorEO, in the cloud, by default: Claude Sonnet 5.5, through the server of MonitorEO. There is nothing to install or download, but it needs a connection. It was chosen in November 2026 by comparing several models on typical questions of the assistant: it got almost all of them right and it is fast. Each user can ask it up to 400 times a day (a message takes from 1 to 8 requests).
• A model on this computer (Ollama): answers worse: an open model that runs on your computer, for those who prefer that the conversation does not leave it or who work offline. The models that fit in a computer are much smaller and make more mistakes: they choose the wrong variable, place or dates, or ask permission for what they could do. If you use one, always check How it is computed in each result.

The header of the window says which model answers. The models of the cloud also get longer instructions: they give step-by-step explanations of how to use the assistant and of the concepts of remote sensing, with more sections of this manual.

The engine of the models on your computer (Ollama)

Only if you choose a model on your computer. The assistant uses Ollama, an open program that runs language models on the computer:

• If you already have Ollama installed (its application, open or not), it uses that one and its models.
• If not, Install downloads the official version of Ollama from GitHub, checks that it is the published one (its SHA-256) and saves it in ~/MonitorEO/ollama, without administrator rights and without installing anything in the system. The assistant starts it when it opens and stops it when it closes.

The models are saved where Ollama saves them (~/.ollama/models), so an Ollama application that you install later can use them too. The Ollama version is fixed by MonitorEO and is updated with it.

The models on your computer

With A model on this computer (Ollama), the Model on this computer drop-down has the three recommended models, with their size; qwen3:8b (recommended) is the first. Those already on your computer have ✓ downloaded, and below it says that nothing needs to be downloaded. If the chosen model is not downloaded but another one of the list is, the assistant uses that one when it opens. For another Ollama model that uses tools, type its name in the box next to it and click Download (or Install, the first time).

Model	Download	Comment
qwen3:8b	5 GB	Default. Tested: about 4 s per answer on a Mac M3 Pro with 18 GB (the first one, about 20 s, while the model loads)
qwen3:4b	2.5 GB	Faster and with less memory; makes more mistakes (untested)
gpt-oss:20b	14 GB	For computers with 16 GB or more of free memory (untested)

A computer with 16 GB of memory is advisable for qwen3:8b.

Ollama on a server

If Ollama runs on another computer (for example, a server of your organization, so as not to download the models on every computer), change "ollama": "auto" in ~/MonitorEO/assistant.json to its address, e.g. "ollama": "http://servidor:11434". The text of the conversation then travels to that server; the computations are still done in Earth Engine, with the shared account or with yours.

Always the latest version

Like monitoreo mcp, monitoreo assistant checks at start whether there is a newer version on GitHub and, if there is, installs it and starts with it; meanwhile (a minute) a small window Updating MonitorEO… is shown. Without a connection, it starts with the installed one.

While it is open, it also checks every 6 hours: if a new version comes out, the button Update MonitorEO appears at the top. When you click it, it installs the version and the assistant restarts by itself (in a minute: its window closes and opens again). The conversation is lost, but not the results, the layers or the reports, which are in the MonitorEO folder.

monitoreo assistant --no-update     # without checking the version
monitoreo assistant --port 8800     # on another port (default, 8765)
monitoreo assistant --browser       # in the browser, instead of in its own window
monitoreo assistant --no-browser    # without opening the window or the browser (server only)

If it is already open, monitoreo assistant (or the double click on the Desktop) does not start another one: it shows the one that is open.

Use

The window has two parts: on the left, the conversation; on the right, the results, with the map at the top. In a narrow window, the two parts are in tabs (Conversation and Results).

Type in the box at the bottom (Enter to send, Shift+Enter for another line) or click one of the examples. More examples:

• "Which sensor suits me to follow the vegetation of a small farm since 2015?" (advice, no computation)
• "Annual NDVI of Sentinel-2 in Sierra Nevada National Park from 2019 to 2024"
• "Compare the summers of 2023 to 2025 with those of 2018 to 2022 in Doñana with the NDVI"
• "Map of the median NDVI of summer 2024 in the study area parcelas"
• "Download as GeoTIFF the median Sentinel-2 NDVI of summer 2024 in the study area parcelas" (to open it in QGIS or another GIS)
• "Export to my Google Drive the Sentinel-2 NDVI of 2023 of Sierra Nevada Natural Park" and, later, "Has the export finished?" (for large regions at full resolution)
• "Annual trend of the MODIS land surface temperature from 2005 to 2024 in Cazorla Natural Park"

The more specific the request (variable, sensor, place and dates), the better. If something is missing, the assistant suggests an option and asks.

Questions about MonitorEO

It also answers questions about the tool with this manual, which it downloads from GitHub (and saves in MonitorEO/wiki to use it offline): "What is the NDVI?", "How often are there Sentinel-2 images and what is their resolution?", "Which cloud filters are applied?", "How do I cite MonitorEO?". Below the answer, the sections of the manual it took it from appear (From the manual: …), with their link.

If you ask for something the assistant does not do, it tells you which MonitorEO tool does it, with its manual: computations that repeat by themselves or of many regions at once, the Python package or the command line; editing layers or making maps to print, the QGIS plugin; working in R, the R package; using it without installing anything, the Earth Engine app; publishing reports, the results viewer.

Protected areas

The assistant looks for the area by its name in the World Database on Protected Areas (WDPA) and shows you the list: the same place often has several designations with different boundaries (National Park, Natural Park, SAC, SPA, Ramsar site...). The one that matches what you asked for comes first and highlighted, and it is the one used. To compute with another, click it: the assistant repeats the computation with it. If you click it while it is computing, it does so when it finishes.

If the model gets the identifier wrong (it invents one or copies one), the assistant looks up the place you named by itself and, if only one matches the designation you asked for ("Sierra Nevada National Park"), computes with that one; if not, it shows you the list for you to choose.

Your study area

To compute on your polygons (farms, plots, management zones...), click Upload study area or drag the file onto the page. A GeoPackage, a GeoJSON, a KML or a shapefile compressed in a .zip are valid, in any coordinate system. The assistant checks the layer and tells you how many polygons it has and their area; then it appears under the header, in Study areas, with its name (the × removes it), and you can ask for computations on it: "Annual NDVI in the study area parcelas since 2018".

Each polygon gives its own series or table row. If the layer has a field with a different name for each polygon (nombre, name, id...), it is used to label them. The study areas are copied to ~/MonitorEO/capas and are still there next time.

Results

Each result comes in a card: the chart or the map (with its legend), the table (its first 6 rows; See the N rows shows them all), See on the map, How it is computed (region, variable, sensor, period, metrics, filters and method) and Download CSV or Download TIF. The GeoTIFFs (float32, NoData -9999, in the UTM projection of the region) are also saved in the MonitorEO folder, with a preview of the map in the card and in the report. Large regions are downloaded with larger pixels (ask for it: "at 100 m") or exported to Google Drive. Check How it is computed before using a result: it is what has really been computed, whatever the model says.

Maps

The maps served by Earth Engine are shown large at the top of the results, with zoom and pan: that of a map you asked for (of a variable, of a trend or of the comparison of two periods) and, with each series, statistics or comparison on a region, the map of the median of the variable in its period, which the assistant asks Earth Engine for when the result is finished (it takes a few seconds). If the period is longer than a year, this map is that of its last year (the median of several years of images takes minutes to draw), with the fixed color scale of the variable; that of some statistics is the map they were computed from. With a trend (Mann-Kendall and Sen), the map is the one of the trend in each pixel: Sen's slope per year (green, rising; red, falling), with a fixed scale of ±2 % of the range of the variable per year. A trend on an area always comes complete, even if you ask only for the one of each pixel: the table (Sen's slope, Kendall's τ and p-value of each month or part of the year, for the whole area), its chart and the map of each pixel. If there are several maps, choose one in Map (each with the title of its result, which includes the area) or with See on the map in the card.

In Date you choose the map of each year, month or 16-day period of the result: the steps of a series (monthly: each month; annual: each year, or each summer if a season was asked for) or, in a map of several years, each year. The first in the list is that of the result: the interannual value if asked for (for example, the mean of the annual medians), the whole period or its last year. Each date is prepared in Earth Engine when you choose it (a few seconds) and uses the same color scale as the map of the result, so that they can be compared. The value of a pixel is that of the date you see, and Add this view to the report saves that date. In a trend of each month (or each 16 days), in Part of the year you choose the annual trend (that of the values of each year, the first) or that of each month. In the trend maps, the Significant pixels only (p < 0.05) box leaves only the pixels whose trend is significant according to the Mann-Kendall test; it is kept when you change month and can be unchecked at any time. The view added to the report is the one you see, with the box checked or not.

You also choose:

• Base map: Map (OpenStreetMap); Sentinel-2 cloudless (EOX, whole world), the EOX Sentinel-2 mosaic of 2023, which serves as an orthophoto outside Spain (up to about 10 m pixel; license CC BY-NC-SA 4.0, for non-commercial use, with its credit, which goes on the map and in the report); or PNOA orthophoto (Spain only), from the IGN, more detailed.
• Opacity of the Earth Engine map over the background.

To query a pixel, click Query pixel (the cursor becomes an arrow) and click on the map: you get the value of the pixel in that map (in a trend map, its Sen's slope and its p-value) and the time series of that pixel (with the step of the series of the result; in a map, 16-day, monthly or annual according to the period), computed in Google Earth Engine with your account. Click the button again to finish and move the map without querying. Keep the series as a result adds it to the results, for the report.

Add this view to the report saves what you see (the framing, the background, the opacity and the marked pixel) as one more result, with the legend and the credits of the background. The Earth Engine maps expire after a few hours; the assistant asks for them again by itself.

Report

Create report (N) gathers the results marked In the report (also the map views and the series of a pixel that you saved): it asks for a title, saves it in ~/MonitorEO/informes and opens it in the browser. It is a single HTML file, with the charts and maps inside, that can be moved, sent or opened offline. For the PDF: Print or save as PDF and, as destination, Save as PDF (turn off "Headers and footers" in the options of the browser's dialog).

Each result has its sheet (how it was computed), the chart or the map and the Python call that reproduces it. The table of values is only included if you check Table in the report in its card (the long ones, in an appendix at the end); the values are always in the CSV. At the end come the citation of MonitorEO, the funding and the logos. The report has no text from the model and no interpretation of the results.

New request starts another conversation. The files of the results stay in ~/MonitorEO.

With --browser, if you close the tab or the browser while it computes, the computation goes on: when you open the assistant again, the page shows the whole conversation, with what is running and the results that have arrived meanwhile. In its own window, closing it closes the assistant (if it is computing, it asks first).

Limits

• The model can make mistakes, above all those on your computer, which are small: it misreads a list, asks permission for something it could do, names the area badly in its answer or goes on too long. That is why the results are checked in their sheet, and the app does not compute with a protected area identifier that did not come from the search (or that you did not type yourself): it looks up the place you named. Links to the manual that do not exist are replaced by the section that answers the question.
• The exports to Google Drive go to the Drive of your Earth Engine account, so they need your account and your project (not possible with the shared account): the assistant says when they finish, but the files are opened from Drive. Publishing in the results viewer is only for its administrators, with the MCP or Python.
• The model of MonitorEO needs a connection and has a limit of requests a day for each user; offline, or if you reach it, use a model on your computer.
• One query at a time. Long series over large areas take minutes: ask for a monthly or annual step.
• Only on your computer: the assistant only answers on 127.0.0.1 and with the key in the address that opens its window.

If something fails, see Troubleshooting.