# DeepSeek

DeepSeek's complete REST surface — the model list and the prepaid credit balance.

DeepSeek publishes an inference API and almost nothing else. This plugin covers the whole of the rest of it, which is two endpoints.

## What you can manage

- **Models** — the models this key can call, with the documented per-model concurrency cap. Read-only.
- **Balance** — the account's prepaid credit, one row per currency, split into granted and topped-up amounts. Read-only.

## Credentials

One field. [platform.deepseek.com](https://platform.deepseek.com/api_keys) → **API keys**. The value starts with `sk-`.

There is no separate admin or management key, and no API for creating or revoking keys, so rotation happens in the DeepSeek console.

![DeepSeek Add-account form with the single API key field](https://agent-assets.infrawrench.com/docs-screenshots/plugins/deepseek/add-account.png)

## Tips & limits

- **There is no speech API.** DeepSeek ships no text-to-speech or transcription endpoint, so this plugin has no [Speech tab](../features/speech-testing.md).
- **There is no usage or billing API either** — only the point-in-time balance. A cost chart built from a balance snapshot would be a fabrication, so this plugin declares no cost capability at all and shows the balance as a balance.
- **Rate limiting is by concurrency, not by requests or tokens per minute.** DeepSeek caps concurrent in-flight requests per model and answers `429` over the cap. Those caps are published in the docs rather than returned by the API, so the model page fills them in.
- **Balance amounts arrive as decimal strings**, not numbers — worth knowing if you script against the same endpoint.
- **The canonical base URL has no `/v1` segment.** `/chat/completions`, `/models` and `/user/balance` are the real paths; `/v1` is accepted only so the OpenAI SDK can be pointed at DeepSeek unchanged.