Where things stand: configuration yes, assistant not yet
This is worth saying up front, because the decision depends on it. In the platform’s credentials module, the “LLM” type has four providers: GPT (OpenAI), Gemini (Google), Claude (Anthropic) and DeepSeek. Each of them takes the same set of fields:
- API key - a secret field, required,
- model - required, entered as text rather than picked from a closed list,
- base URL - optional, for overriding the provider’s default endpoint.
The admin assistant, however, has adapters for three providers in the platform’s public code: Anthropic, Google and OpenAI. If you point its settings at a configuration with DeepSeek as the provider, the assistant reports a “not configured” state and executes nothing. That is deliberate: given credentials it cannot use, the assistant fails closed instead of trying another provider.
In practice, you can already store a DeepSeek configuration safely and reference it from settings, but the assistant will only start using it once a DeepSeek adapter exists for it.
What happens to a DeepSeek key inside the platform
You create a configuration once, give it a name and a code, and settings refer to it by that code. Rotating the key happens in one place.
- The key is encrypted at rest and never comes back in plaintext: after saving, you only see that the field is set.
- Creating, editing and deleting a configuration are written to the audit log, with secret fields redacted.
- The type and provider are locked after saving. Moving from DeepSeek to another provider means a new configuration, not an edit of the old one.
- A configuration that a setting still refers to cannot be deleted.
What DeepSeek provides on its side
DeepSeek is a language model provider based in China (Hangzhou DeepSeek Artificial Intelligence Co., Ltd.). Its API documentation establishes a few things that matter for an integration like this:
- The API uses a format compatible with the OpenAI and Anthropic APIs, and the key is generated in DeepSeek’s developer platform.
- The models support tool calling. That is a hard requirement, because the assistant does not write database queries: it picks operations from a closed catalogue of tools registered in code. Strict adherence to a tool’s schema is a beta feature at DeepSeek, served from a separate base URL.
- Model names change: the documentation describes older names that are still accepted but are now served by a newer model. This is why the model name is a configuration field in the platform and not a value fixed in code.
- When the concurrency limit is exceeded, the API returns a 429 error. Under heavy load the connection is kept alive, and if inference has not started within 10 minutes the server closes it.
Data location is a separate matter. DeepSeek’s privacy policy names China as the place where data is processed and stored. What reaches the model is the text of the instruction plus the search results the assistant builds its plan on, for example product names, warehouse names or order data. Before choosing this provider, check DeepSeek’s current terms for the API and agree it with whoever is responsible for data protection in your company.
How the assistant behind this list works
The assistant reads an instruction typed into the command palette in plain language, in Polish or English, and works out a plan against the platform’s own data model. It shows that plan before doing anything and only executes after you confirm. Every action stays inside your permissions and lands in the audit log, together with the instruction text, the provider and the model.
A few facts that do not depend on the provider:
- a plan that is not confirmed within 10 minutes expires,
- by default a single instruction can affect at most 500 records, and a larger plan is blocked at the preview stage,
- a configuration change takes effect from the next instruction, with no restart.
What to choose today
If the assistant needs to work right away, point its settings at a configuration for one of the three supported providers: Claude, Gemini or GPT. The range of instructions is the same; only the model answering differs.
If DeepSeek is a requirement in your organisation, for instance for vendor diversification, the missing piece is an adapter on the assistant’s side. That is a change in the assistant module’s code, not in settings. To agree the scope and timing of that work for your deployment, get in touch.