Your data

Last checked against the product on 12 October 2026

Aevia works with your own openLCA database, so the first question is what happens to the data in it. In short:

• Your database stays in your openLCA. Aevia never imports it, hosts it or copies it as a whole.
• Calculations on it run in your openLCA, under your licence, and are never charged for or counted against your limits.
• Aevia keeps the structure of your product systems and every result the assistant calculates. A dataset the assistant opens stays in the chat only while the assistant works with it; after that the chat keeps its name and identifier, not its contents.
• The AI model running the conversation reads what the assistant reads. In Aevia's own chat that model belongs to a provider Aevia sends the conversation to; from Claude or ChatGPT, it is your own assistant.

The rest of this page says exactly what, where, and for how long. If something here does not match what you see, write to hello@aeviamodeler.ai; this page is only worth anything if it is exact.

Three ways your data can travel

In Aevia's chat, with your own openLCA

1. Your openLCAon your computer
2. connector app, encrypted
3. Aevia's serverGermany
4. the text the assistant read
5. AI model providerAnthropic, OpenAI or DeepSeek

From Claude, ChatGPT or another assistant (MCP)

1. Your openLCAon your computer, if connected
2. connector app, encrypted
3. Aevia's serverGermany, no AI model
4. tool results
5. Your assistantunder your own account

On a demo database

1. Demo databasepublic data, hosted by Aevia
2. inside Aevia
3. Aevia's serverGermany
4. as above
5. Model provider or your assistant

In Aevia's chat, step by step

1. You ask, and the assistant picks a tool — search the database, open a process, build a system, calculate. The request goes from Aevia's server to your computer.
2. The connector app carries it. The small program on your computer dials out to Aevia over an encrypted connection; it opens no port on your machine, and it never writes the content of requests or answers to its own logs.
3. Your openLCA answers. A search returns names, categories and locations. Opening a process returns the whole dataset: every exchange with its amount, its description and its documentation. A calculation runs on your computer and returns its results.
4. The answer passes through Aevia's server on its way to the assistant. The server sees it, in memory; it writes to storage only what the table below lists. To keep searching fast, it also holds the list of names in your database (processes, flows, methods and the like, without their contents) in memory while it runs, refreshed when it is older than 30 minutes, and never written to storage.
5. The model provider receives the text of what the assistant read, with Aevia's internal identifiers removed. For a search, the names. For an opened process, its exchanges — flow names, amounts and units, the first 20 on each side unless the assistant asks for more, up to 100 at a time — and its description, location, validity period and data-quality entry; on a later turn, only if the assistant opens it again. For a calculation, the impact totals and the main contributing processes with their values.
6. The chat is saved. A process the assistant opened is saved in full while the assistant is still working with it, so that it can come back to it after you confirm a step or a calculation finishes. When you send your next message, or about 20 minutes after the reply (24 hours if the chat is waiting on a question or form you have not answered), only its name, identifier, location, reference flow and number of exchanges stay: not the exchanges, their amounts or its description. Its card then reads it again from your openLCA when you open it, and on a later turn the assistant opens it again if it needs the figures. What the assistant wrote in its replies, and the inputs it gave other tools (an authored process or a composed system can carry amounts it copied), stay as written. Every result the assistant calculates is saved as an assessment, in your workspace or, in a chat with none, in a hidden one that belongs to the chat.

If instead of the connector app you set up a direct network connection to an openLCA server, the same applies; Aevia accepts one only if its address starts with https, so that link is encrypted too.

While Aevia is in early access it also keeps, for 30 days, a complete record of each request the assistant makes to a model — what was sent, tool results included, and what came back — to investigate problems. Aevia's administrators can read these records, and each time one reads another person's, that is logged.

From Claude, ChatGPT or another assistant

Through the MCP Connector, Aevia sends nothing of yours to a language model, except that a knowledge-library search goes to OpenAI to be matched. Your assistant decides what to call; Aevia carries the request to your openLCA, by the same encrypted connection, and returns the result to your assistant, which handles it under your account and your agreement with its provider — which may be your organisation's. Aevia does not keep the conversation you have there.

What Aevia keeps from this route:

• whatever your assistant asks it to save — product systems, assessments, comparisons — in your workspace, like any other content;
• a record of the connection: its name, the access level you approved, and when it was last used;
• the name and identifier of each dataset the assistant found or opened, so it can refer to it again: until your account is deleted;
• which assistant software connected and which of Aevia's playbooks it loaded: 180 days;
• for each tool call, which tool, whether it succeeded and how long it took, without its contents or its result: 180 days;
• the text of each knowledge-library search and whether it found anything: 30 days;
• the result of a call that took too long to return at once, for your assistant to collect: 24 hours.

On a demo database

The shared demo databases hold public data that Aevia's operators loaded. Calculations on them run on Aevia's server and, when they reach the engine, count toward your plan's daily calculation limit. None of your own data is in them, so a process opened from one stays in the chat in full, like the rest of the conversation. Results from a demo database can be precomputed or reused between calculations; results from your own engine never are.

What Aevia keeps

The full list, with everything Aevia keeps about your account and your payments, is in the Privacy Policy.

Which AI providers, and on what terms

In Aevia's chat, the conversation goes to the provider of the model it runs on. The chat shows which model that is, and you can change it. The providers' terms differ, and the difference matters most for licensed data:

• Anthropic (Claude). Its commercial terms prohibit training models on content sent through its API. It deletes inputs and outputs within 30 days, except where it keeps them longer to enforce its usage policy or meet legal requirements. A few background tasks use an Anthropic model: reading a document you add to your library, drafting a workspace's goal, and checking a report's proposed corrections.
• OpenAI (GPT). OpenAI states that data sent to its API is not used to train its models unless the customer opts in, and keeps it for up to 30 days to monitor abuse. Aevia asks it not to store the conversation itself. OpenAI also matches knowledge-library searches.
• DeepSeek, offered in the model picker but not the default on any plan. DeepSeek processes and stores data in the People's Republic of China, and its privacy policy lists training and improving its models among the uses of the content it receives, without an exemption for API customers. Do not choose a DeepSeek model for a conversation about data your licence does not let you share on those terms.

If you add your own API key for a provider, your conversations on its models go through your account, under your own agreement with it. When the assistant searches the web, the query goes to Tavily, a search provider in the United States, whose terms let it keep query data and use it to improve its services.

What Aevia does not do

• Import or host your database. Aevia's own catalogue holds only the public demo databases its operators load.
• Put results from your engine into a shared cache or a precomputed matrix.
• Charge for calculations on your engine, or count them against your limits.
• Open a port on your computer: the connector app only dials out.
• Write the content of requests or answers to the connector app's logs.
• Share anything in your workspace with other users, or sell it.

What Aevia writes into your openLCA

On your own openLCA connection, unless you mark it read-only in its settings, the assistant can create unit processes, flows and product systems in your database, and delete what it created. In Aevia's chat each one is shown to you first, and made only when you press its button. From another assistant, a call without confirmation returns a preview and one with confirmation writes; whether you are asked in between depends on that assistant. What it writes goes in a folder named Aevia/ followed by the workspace's name (or Aevia/Unfiled), is tagged aevia, and says that Aevia wrote it. A product system that openLCA links itself, with its own "Create product system", is placed where openLCA puts it, at the top of the database.

Some calculations also need a temporary product system in your database. For a system built in Aevia, or recalculated from Aevia's saved copy, Aevia writes one in Aevia/Temporary. When openLCA's own linking builds one, openLCA places it at the top of the database, named after its process. Aevia deletes these when the calculation ends, and keeps a list so it can remove any that a dropped connection left behind; if that list cannot be saved, a temporary system can stay until you delete it.

The shared demo databases are never written to.

Your licence

Your licence terms with your data provider govern how you may use their data, and Aevia does not change them. Aevia is not affiliated with or endorsed by any database publisher, and no publisher has reviewed or approved this page. If your licence restricts sharing dataset content with third parties, the routes above show which steps a dataset crosses, so you can judge. The MCP route keeps the AI model under your own account, and possibly your organisation's agreement with its provider.

Deleting things

• A chat — its history goes, including what it kept of the processes opened in it, and with it the records of its model requests.
• A product system — its assessments go with it. Contribution and sensitivity analyses made from it keep their own copy of the results until you delete them.
• A workspace — everything in it.
• Your account — from Settings → Profile → Delete account. It takes effect when you confirm the link we email you, ends any subscription at the same moment, and deletes everything above. Anything Aevia wrote into your own openLCA stays there, because that database is yours. Encrypted backups roll off within 90 days.