The Promise
- →Uploads are deleted from our servers after parsing. The parsed markdown lives in
.omenon your machine. We keep no copy. - →Inference routes to Anthropic, OpenAI, or open-weight models like Qwen, DeepSeek, and Kimi, picked per query for quality.
- →Every provider in our routing operates under commercial or zero-retention terms. None of them train on your data.
- →Need stricter? Self-host an open-weight model, point us at it, and nothing leaves your infrastructure.
Your files never stay with us
When you upload a schematic or datasheet through the web app or CLI, the file lands in our file system temporarily so we can process it. We parse it into structured markdown, send the parsed version back to your device, and then delete both the original upload and the parsed copy from our infrastructure.
Everything ends up in a .omen folder inside your project, locally. That folder is the source of truth from then on. When the agent needs context from a datasheet you uploaded last week, it reads the markdown out of .omen on your machine and includes it in the request. We do not keep a copy.
This means if you delete .omen, the data is gone. There is no copy on our side to recover from, and that is the point.
How AI inference works
The agent itself runs locally on your machine. The piece that needs to talk to a remote service is the model call, the actual LLM inference, and we are deliberate about how that is routed.
We pick the best model per query. For some tasks that means a frontier model from Anthropic or OpenAI. For others it means an open-weight model like Qwen, DeepSeek, Kimi, or GLM. Either way, the inference path is held to the same standard.
When we route to Anthropic or OpenAI, we use their commercial APIs under terms that contractually prohibit training on your data. Logs are retained briefly for abuse monitoring and then deleted.
When we route to an open-weight model, we serve it through inference providers that operate under zero-retention agreements. They do not train on your data and they do not store it. The model is open, the host is contractually bound, and your firmware context is gone the moment the request completes.
For users with stricter requirements, you can also bypass our routing entirely and point us at your own self-hosted inference endpoint. If you are working in defense, medical, ITAR-regulated, or otherwise export-controlled contexts, this means the request never leaves your infrastructure. The model weights are open, the host is yours, and we never see the traffic.
