NextCore Agent

Connect your API keys for Anthropic, OpenAI, Gemini and a dozen other providers, order them into a fallback chain, and see models, prices, quotas and recent calls.

Sidebar › NEXTCORE AGENT. This is where you connect AI providers for the whole app. Every other AI feature calls through here: keyword suggestions and classification in NEXTCORE GROUPCRAWL, post spinning in NEXTCORE AUTOMATIONS, the Director, voices and the fallback Cloud API (images and video when the profile pool runs out) in NEXTCORE MEDIA, and script rewriting in NEXTCORE TOOLS.

The app calls providers directly with your keys, not through the NextCore server. Model usage is billed to your provider account; NextCore charges nothing extra and does not issue keys.

The screen opens straight onto the provider list; the line under the title shows the number of providers, how many are on and the number of models. The Chat button at the top of the page opens the assistant that controls profiles (see Trợ lý AI (AI assistant)); click Providers in the same place to go back.

Since 0.3.171, the Chat screen has a single header strip named Agent: the secondary line shows the running task (description, number of profiles, start time), and the Task history and ← Providers buttons sit at the end of the button row. In windows narrower than 900 px, the Awaiting your approval block moves above the board and the button row moves below the title, so anything you need to click stays in view.

#Sections

#Add a provider

1. Click Add provider. The catalog is split into groups:
   • Đám mây (Cloud): Anthropic, OpenAI, Google Gemini, Groq, Mistral, DeepSeek, xAI (Grok), Moonshot (Kimi), Cerebras, Together AI, Fireworks AI, NVIDIA NIM, Z.ai (GLM), MiniMax.
   • Voice only: ElevenLabs. This key is used for virtual characters' voices and does not join the chat fallback chain.
   • Cổng gộp · một khoá dùng nhiều hãng (Aggregator · one key for many vendors): OpenRouter, Hugging Face.
   • Chạy trên máy này · miễn phí (Runs on this machine · free): Ollama, LM Studio, llama.cpp server, vLLM, Custom (OpenAI-compatible).
2. Choose an entry. Fill in the Key label (default «Main key») and API key. For entries that run on this machine, you can edit the Server address and the key is optional.
3. Click Check and save. The app requests the provider's model list (no tokens used). If the check passes, the card shows Key OK, the number of models and the response time; choose the Default model right in the dialog.

If the check fails, the error appears under the field and the app has Nothing saved yet.

With no providers yet, the empty screen has a Connect Ollama on this machine button: install Ollama from ollama.com and click it, with no key and no charge.

The catalog also has a Not supported (technical reason) section: Azure OpenAI, AWS Bedrock, Google Vertex AI, Cohere and Perplexity, each with its reason.

#Voices

Since 0.3.156, virtual characters' voices can be chosen by source: ElevenLabs (its own key, in the Voice only group), MiniMax (reuses the MiniMax chat key you have connected), OpenAI, or a TTS server you host yourself. The voice field shows voice names instead of IDs. When ElevenLabs reports the monthly character quota used up (429) or no credit left, the card says so plainly rather than reporting a bad key.

#Where keys are stored

Keys are stored on this machine. On Windows and on Linux with a keyring, keys are encrypted with the operating system's keyring. On macOS (and Linux without a keyring), keys are currently stored unencrypted in the app's data folder, and the card says so. The screen only shows a masked version.

#Fallback chain

Drag the handle on the left of a card to change the order (or Alt + ↑ / ↓). The first provider in the chain that is currently serving gets a Serving chip. When a call fails, the app acts according to the type of error:

If a provider has several keys, the next key is used when the main key hits its rate limit: expand the card and paste another key with a label such as «Backup key».

When the whole chain is exhausted (or there are no keys yet) and Muse via profile is on, the NEXTCORE MEDIA Director replies through Muse: free, about 25 seconds slower. Muse needs a profile signed in to Muse AI in the NEXTCORE MEDIA › Sources pool. Group scoring and suggestions in NEXTCORE GROUPCRAWL and the browser-controlling assistant only use API keys and never fall back to Muse.

#Expanded provider card

• API key: each key with its state, Key OK, Rejected, Paused until …, Not checked, and response time; add or remove keys.
• Available models: a table with MODEL, CONTEXT, IMAGES and USD / 1M IN · OUT. Click a model to make it the default. Prices come from the provider's API when available; for providers without a pricing API, public reference prices are used and the source is stated.
• Quota: remaining requests and tokens per minute, shown after the first call (read from the response). DeepSeek and Moonshot show the remaining balance.
• Connection: the server address (for entries that run on this machine).
• Remove provider.

#Spending cap and approval gate

The Daily spending cap is set in SETTINGS › Limits (click the Today … USD chip to open SETTINGS). The same section has Max profiles running in parallel, the Always pause for approval switch and Payments always need approval (cannot be turned off). When the cap is reached, all AI work stops until 00:00 or until you raise the cap.

API keys for every provider live in one shared store; review them in SETTINGS › AI assistant and providers (click a provider to see its keys).

When the assistant controls a profile, posting, commenting, messaging and friend requests pause for your approval; payments always wait for approval. Details on the approval gate, per-task caps and how to read errors: Trợ lý AI (AI assistant).

#Common errors

• Key check fails, «Nothing saved yet»: the key is wrong, expired, or API access is not enabled. Create a new key on the provider's site and check again.
• Chip «bỏ qua: hết số dư» (skipped: no balance left) or «hết tiền / chạm trần chi tiêu» (out of credit / spending cap reached): top up with the provider; the app puts it back into the chain automatically from the next day.
• Mọi khoá đang vượt hạn mức (429) — thử lại sau N giây (Every key is over its rate limit (429), try again in N seconds): wait the number of seconds shown, or add another key or provider to the chain.
• «<Nhà cung cấp> tạm ngắt tới <giờ>» (« paused until »): the spending cap was reached or the balance ran out. The app reopens it automatically at that time without reconnecting; to use it right away, raise the cap in SETTINGS › Limits.
• «Chưa chọn model cho …» (No model selected for …): open the provider card and click a model under Available models.
• «Chưa kết nối nhà cung cấp AI ở NEXTCORE AGENT» (No AI provider connected in NEXTCORE AGENT): no provider is on. Click Add provider.
• Model hiện sau khi kiểm khoá (Model appears after the key is checked): the card has no model list yet. Click Test.
• Ollama cannot be checked: Ollama is not running, or its address is not http://localhost:11434/v1. Open Ollama and click Test.