Ask, don't click: why energy data needs AI agents rather than another dashboard

Almost every start-up in energy begins with a dashboard. Maps, charts, filters and an export button. Anyone who looks after twenty parks already has a control room, spreadsheets, reports from the operations manager and a full inbox. What they want are answers, asked for the way you would ask an experienced colleague.

In this article, we explain why we are opening our platform to AI agents. We describe what the Model Context Protocol (MCP) is, what our MCP server can do and what it cannot do yet. One thing first: the server is in development and not yet generally available. The description reflects the state of development on 1 November 2026.

The problem with the next dashboard

Every tool has its own logic: its own login, its own menus, its own filters, its own export formats. For a portfolio manager, by contrast, the question is simple. Which of my parks are in a flood zone? Which assets are still missing from our overview? What does the park we are currently reviewing look like? In a dashboard, you find the answer if you know where to click. If you open the tool once a quarter, you usually no longer remember.

There is also a simple connection: a platform only creates value if people use it. A dashboard nobody opens answers no question. That is not an argument against dashboards. Maps, loss waterfalls and uncertainty bands remain important when you want to check a result or explain it to your committee. But the entry point is shifting. First comes the question, then the chart that backs up the answer.

What Gartner expects and what it warns about

The research firm Gartner describes this shift too. In a press release of 26 August 2025, Gartner predicted that up to 40 percent of enterprise applications would include integrated, task-specific AI agents in 2026. According to Gartner, the figure was less than five percent in 2025. By 2028, according to Gartner, a third of user experiences will shift from native applications to agentic front ends. In other words, users state their goal and an agent works its way through the applications, instead of the person operating each application individually.

The same source is also sober. In a press release of 25 June 2025, Gartner expects more than 40 percent of agentic AI projects to be cancelled by the end of 2027. Gartner names escalating costs, unclear business value and inadequate risk controls as reasons. Gartner also warns about "agent washing": existing products such as assistants, automation software or chatbots are marketed as agents without substantial agentic capabilities. Gartner estimates that only about 130 of the thousands of vendors in this field are real. Its recommendation is to use agentic AI only where it delivers clear value.

We draw one conclusion from this: not every product needs its own chatbot. We don't want to sell you another assistant that you have to learn. We want to make our data accessible to the agents you already use or will soon use.

What an agent is and what a good colleague can do

An AI agent is a system based on a language model that does not just write text but uses tools. It retrieves data, checks it, carries out steps and reports what it has done. The comparison with a colleague helps to understand the requirements. A good colleague knows where the data is. They ask when something is unclear. They say when they don't know something. And they don't delete your documents without asking.

That is exactly the standard for an agent in energy. Without access to reliable data, a language model guesses. Without clear limits, it can do damage. An agent is therefore only as good as the interfaces it works through: which data it may see, which steps it may carry out and which it may not.

The Model Context Protocol: a shared plug

This is where the Model Context Protocol comes in. Anthropic introduced MCP on 25 November 2024 as an open standard for connecting AI assistants to the systems where data lives. Instead of building a separate integration for each assistant, a data provider offers one MCP server. Any assistant that supports MCP can use it.

On 9 December 2025, Anthropic donated MCP to the Agentic AI Foundation, a directed fund under the Linux Foundation co-founded by Anthropic, Block and OpenAI. According to Anthropic, there were more than 10,000 active public MCP servers at that point. ChatGPT, Cursor, Gemini, Microsoft Copilot and Visual Studio Code are among the products that have adopted MCP. For you, this means: you choose the assistant, we provide the data. Which features your assistant supports in detail, and whether it may connect to external servers, is something to clarify with your IT department.

What our platform calculates

An agent is only as useful as the data behind it. Our platform structures a portfolio as plant, asset and component, for example a turbine, a PV component or a shared grid connection. For wind, we calculate yield per turbine from hourly site weather and the power curve. The catalogue covers about 900 turbine models. The result is a loss waterfall with P90/P50/P10 bands. For solar, we use a pvlib chain with hourly irradiance and temperature. The basis is hourly weather data for the last 20 complete years. For the outlook, we use EURO-CORDEX climate projections on a grid of about 25 kilometres.

On top of that comes a risk and measures register. Risks such as icing, lightning or an ageing rotor are linked to measures, which we value in euros at your price, your investment horizon and your discount rate. For Germany, data from the market master data register (MaStR), substations as connection points and a congestion layer are added. Weather and climate projections cover more than 30 European countries.

What our MCP server can do

The agent signs in with the customer's API key. Every tool works exclusively on the account that owns this key. No tool accepts another account as a parameter.

Data in. The agent can upload files, which are stored in the account as received: CSV, Excel, XML, JSON, PDF or ZIP, up to 7 MB through this route. It can have a portfolio description checked without writing anything. The check lists every problem found. Only then does it import. The import is additive: a plant or asset whose name already exists stays as it is and is skipped. Nothing is deleted or overwritten. If the description has errors, nothing is written at all.

Data out. The agent can export the whole portfolio with plants, assets, turbines and solar components, in the same format the import expects. An export can therefore be edited and imported again. It can also list the existing plants and list and read uploaded files.

Lookups. The agent can search the MaStR live by unit, site, EEG or operator number, or by name, park, operator, postcode and town. And it can find turbines in our catalogue. These are the same tools our AI Importer uses in the app.

An example: you ask your assistant to match the asset list from the operations manager's last email against the MaStR and to create the missing parks in EnviroTrust. The agent reads the list, looks up the units in the register, matches the turbines to the catalogue, has the description checked and reports the problems to you. Then it imports what is missing. Whether your assistant may read your email, however, depends on its other connections, not on us.

Why the import only adds

In the app, our AI Importer proposes an import, and nothing is written until you click "Import". Over MCP, this is different. When your agent imports, that is your own action under your key. There is no second click. That is why we deliberately built the import so that it only adds. An agent that makes a mistake can create a superfluous plant. But it cannot delete or overwrite existing data. That is the rule we would expect of a new colleague, too.

What the server cannot do yet

The first version does not answer yield or risk questions directly. It makes sure that data gets in and out cleanly. That is the foundation, because every analysis starts with a complete portfolio. Results such as yield bands or valued measures remain in the app. Which analyses should come through the interface next is something we want to decide with customers who try the server.

Data protection: you choose the assistant

Our platform runs on servers in Frankfurt. When your agent calls a tool, the results go to the assistant you use, and therefore to its provider. Which provider that is and where it processes data is your decision, governed by your contract. So check with your data protection officer before you connect portfolio data. Treat the API key like a password. Our AI Importer in the app, by the way, runs on Anthropic's Claude, processes data outside the EU and therefore asks for your explicit consent before first use. It is optional.

Limits remain limits

An agent does not make a result more accurate. Our yield bands remain approximate values. The platform is not a yield assessment and not a yield determination under FGW TR 6, does not provide grid operator data and is not financial or investment advice. There is no separate physical yield model for battery storage and grid connections. A language model can, however, phrase a result more confidently than it is. So make sure your agent's answers state their source and limits. A good colleague does that too.

How to get started

You don't need to learn a new tool for this. In a call, we show you your own asset on the platform. If you already use an assistant with MCP support, tell us which one. Then we clarify together whether a test with your portfolio data makes sense.

Sources and historical context

The market figures are Gartner forecasts from 2025, not our own research. The information on MCP comes from Anthropic. The description of the MCP server reflects our state of development on 1 November 2026; features may change before release.

• Gartner: Gartner Predicts 40% of Enterprise Apps Will Feature Task-Specific AI Agents by 2026, 26 August 2025 (task-specific agents, agentic front ends by 2028)
• Gartner: Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027, 25 June 2025 (cancelled projects, agent washing, recommendation)
• Anthropic: Introducing the Model Context Protocol, 25 November 2024
• Anthropic: Donating the Model Context Protocol and establishing the Agentic AI Foundation, 9 December 2025 (Agentic AI Foundation, adoption)