24 · your own endpoint
Running it on Helmcode
DeepSeek Harness speaks the OpenAI protocol, so pointing it at Helmcode is five fields and a key.
What Helmcode is
Open model inference on European infrastructure, behind a single OpenAI-compatible API. Zero logs, flat rate and data that never leaves the European Union.
It is not a model lab and not a consultancy. It is the layer that serves models, operated by us, for teams that want to use open models without building and maintaining the machines.
There is a concrete reason it shows up here: DeepSeek V4 Flash is one of the models we serve. You can run DeepSeek’s harness against a DeepSeek model served from European infrastructure, and neither of the two pieces is closed.
Helmcode pricing
The five fields
In the DeepSeek Harness interface, a custom provider needs an identifier, a base URL, a protocol, a credential and at least one model.
| Field | Value |
|---|---|
| Identifier | helmcode |
| Base URL | https://api.helmcode.com/v1 |
| Protocol | OpenAI-compatible |
| Credential | your key from the console |
| Model | deepseek-v4-flash |
The button that queries the server model list works against our endpoint, so there is no need to type them by hand.
The other models
- Reasoning and agents deepseek-v4-flash Long context, the one that fits this piece
- Volume and code qwen3.6 The best cost to result ratio for repeated tasks
- Multimodal mimo-v2.5 Vision, audio and text
- Efficient gemma4 Lightweight assistants and document work
They are registered on the same provider and you switch between them without touching anything else, because to DeepSeek Harness they are models of the same provider.
It is worth having more than one registered. The model that suits a long reasoning turn is not the one that pays off for classifying a thousand documents.
Three things that will happen to you
- One The identifier cannot be renamed Ever. Requests, saved sessions and credential references all use it. If you get it wrong, you have to create another one and delete the old one.
- Two The key cannot be read back Once saved, the screen only ever returns a redacted version.
- Three Images have to be declared A model you enter by hand is treated as if it did not accept images until you say so. There is no way to ask a server which formats it accepts, so it prefers not to assume.
What this does not change
Nothing the rest of this site says depends on which server you point the harness at. The architecture, the permissions, the session log and plugin composition all work the same with our API, with DeepSeek’s or with a server of yours in a machine at your office.
That is exactly the interesting part of the design, and it is why this section can be so short.