Guide · Updated November 2026

Run DeepSeek inside Cursor, Cline and Continue — with your own key

Every coding assistant worth using now accepts an OpenAI-compatible endpoint. That means you can keep the editor and swap the model behind it: DeepSeek-V4-Flash for reasoning-heavy edits, Kimi-K2.7-Code for long agent runs, Qwen for cheap one-line fixes — all on one balance, paid by card.

The three values every tool asks for:
Base URL: https://api.novatoken.dev/v1
API key: your sk-... key from the dashboard
Model: deepseek-ai/DeepSeek-V4-Flash or moonshotai/Kimi-K2.7-Code

Cursor

  1. Open Settings → Models (Cmd/Ctrl + Shift + J).
  2. Paste your key into the OpenAI API Key field.
  3. Turn on Override OpenAI Base URL and enter https://api.novatoken.dev/v1.
  4. Under Add model, type the model id exactly — e.g. deepseek-ai/DeepSeek-V4-Flash — then enable it and select it in chat.

Note: Cursor's own features (Tab completion, some agent modes) may still use your Cursor plan. The custom key applies to the model you pick in chat.

Cline (VS Code)

  1. Open the Cline panel → gear icon → API Configuration.
  2. API Provider: OpenAI Compatible.
  3. Base URL: https://api.novatoken.dev/v1 · API Key: your sk- key.
  4. Model ID: deepseek-ai/DeepSeek-V4-Flash (or moonshotai/Kimi-K2.7-Code for big refactors).

Roo Code and Kilo Code use the same OpenAI-Compatible provider form — same three values.

Continue

# ~/.continue/config.yaml
models:
  - name: DeepSeek V4 Flash
    provider: openai
    model: deepseek-ai/DeepSeek-V4-Flash
    apiBase: https://api.novatoken.dev/v1
    apiKey: sk-your-novatoken-key
    roles:
      - chat
      - edit
      - apply

Which model for which job

TaskModelWhy
Agentic refactors, multi-file editsKimi-K2.7-CodeTuned for tool use and long repository context (262K)
Reasoning-heavy changes, design questionsDeepSeek-V4-FlashStrong reasoning, 1M context, cheapest per output token
Quick edits, renames, commit messagesQwen2.5-7B-Instruct~2s first token, $0.10 / $0.30 per 1M
“I don't want to think about it”autoEach request routed to a suitable model, billed at that model's own rate

Why agent sessions stay cheap here

Coding agents resend the same context on every turn — system prompt, file contents, previous tool results. On DeepSeek-V4-Flash, input tokens that hit the upstream cache are billed at $0.006 per million instead of $0.30, a 50× cut, applied automatically whenever the provider reports a cache hit. A long session that would cost a dollar of fresh input often bills a fraction of that.

Rough math for a normal day: 200 chat turns averaging 4K input and 400 output tokens is ~800K input + 80K output — about $0.34 on DeepSeek-V4-Flash before cache discounts. Most people top up $5 and forget about it for weeks.

FAQ

Does this work with any OpenAI-compatible tool?

Yes — Cursor, Cline, Roo Code, Kilo Code, Continue, LobeChat, Cherry Studio, Open WebUI, LibreChat, SillyTavern, and anything else that lets you set a base URL. If a tool only accepts a key and no base URL, it is locked to its own backend and cannot use NovaToken.

Streaming and tool calls work?

Yes. SSE streaming, function/tool calling and JSON mode are all supported, which is what agents rely on.

The model list in my editor is empty — what now?

Use the exact ids from the pricing table, e.g. deepseek-ai/DeepSeek-V4-Flash. Some tools have an “add custom model” field; others accept free text. Copy the id character-for-character.

I get 401 — why?

Either the key has a typo, or the base URL is missing /v1, or you pasted the key into the Anthropic field instead of the OpenAI one. All three are easy to hit; check the base URL first.

Is there a minimum spend?

$5, one time. No subscription, and the credits do not expire.

Point your editor at it in two minutes

One key for DeepSeek, Kimi, Qwen and GLM. Card payment, USD, $5 to start.

Buy $5 of credits