NovaToken speaks the OpenAI API. So instead of learning a new SDK, you change two values in whatever you already have open — the base URL and the key. Below are the exact configs for 12 tools, including the ones that hide the setting in a submenu.
https://api.novatoken.dev/v1sk-... key from the dashboardauto to let the gateway pick
| Model id (copy exactly) | Use it for | Price / 1M in · out |
|---|---|---|
deepseek-ai/DeepSeek-V4-Flash | Reasoning, docs, long context (1M). Cached input is $0.006. | $0.30 · $1.20 |
Qwen/Qwen2.5-7B-Instruct | Short chat, classification, extraction — the cheap workhorse. | $0.10 · $0.30 |
zai-org/GLM-5.3-Flash | Everyday writing, bilingual content. | $0.15 · $0.50 |
moonshotai/Kimi-K2.7-Code | Agentic coding, 262K context, multi-file refactors. | $1.20 · $5.00 |
auto | Mixed traffic. Routed per request, billed at the model that answered. | varies |
Cmd/Ctrl + Shift + J).https://api.novatoken.dev/v1.deepseek-ai/DeepSeek-V4-Flash, then pick it in the chat model selector.Cursor's Tab completion and its own agent may keep using your Cursor plan. A custom key applies to the model you select in chat.
All three use the same panel — Settings → API Configuration:
| Field | Value |
|---|---|
| API Provider | OpenAI Compatible |
| Base URL | https://api.novatoken.dev/v1 |
| API Key | sk-... |
| Model ID | deepseek-ai/DeepSeek-V4-Flash |
For large multi-file refactors, switch the model id to moonshotai/Kimi-K2.7-Code. Agents resend context every turn, so DeepSeek's cached-input price ($0.006/M) does a lot of work over a long session.
Edit ~/.continue/config.yaml:
Docker one-liner:
Or in the admin UI: Settings → Connections → OpenAI API → set the URL and key, then add the model ids manually.
Settings → Language Model → OpenAI: set the API Key to sk-... and the Proxy URL / Base URL to https://api.novatoken.dev/v1. Then add the model ids under “Custom models”.
Settings → Model Provider → OpenAI: API Host https://api.novatoken.dev/v1, API Key sk-..., then add models manually with the ids above.
Connection settings → API: OpenAI-compatible → Base URL https://api.novatoken.dev/v1, key sk-..., model zai-org/GLM-5.3-Flash (good bilingual roleplay) or deepseek-ai/DeepSeek-V4-Flash.
| Symptom | Cause |
|---|---|
| 401 invalid key | Base URL missing /v1, key pasted into an Anthropic field, or a trailing space in the key. |
| Empty model dropdown | Most clients do not fetch the catalogue. Type the id by hand, character for character. |
| Requests hang, then time out | A proxy or VPN in between is buffering SSE. Disable buffering, or test with curl first to isolate the network. |
| “Model not found” | Hyphens vs underscores. It is DeepSeek-V4-Flash with hyphens and Qwen2.5-7B-Instruct with a dot. |
Yes, as long as it lets you set a custom base URL. If a tool only accepts a key with no base URL field, it is locked to its own backend.
Yes — SSE streaming, function/tool calling and JSON mode, on every model including auto.
$5, one time. No subscription, credits do not expire.
auto in a coding agent?You can, but agents are exactly the workload where you want a known model — pin moonshotai/Kimi-K2.7-Code or deepseek-ai/DeepSeek-V4-Flash so behaviour stays consistent between turns.
DeepSeek · Qwen · GLM · Kimi. $5 to start, no subscription, billed per token at each model's own rate.
Buy $5 of credits