Official platform API keys
BYOK means using your own vendor account and API key to reach a model. The key is configured locally in Mystilink and usage is billed by the vendor. If you want to use your Mystilink account and balance instead, see Newton LLM; there is no need to add a third-party key here.
Three pieces of information to prepare first
- API key: generated on the developer platform of the vendor that actually serves the API. Do not enter a website password, a Mystilink activation key, or the sign-in credentials of a chat product.
- Model ID: the identifier your current API account is allowed to use. The display name is not necessarily the ID used for requests.
- API access and quota: being able to sign in to a vendor's chat website does not mean the account has usable API quota. Confirm on the developer platform.
Do not assume an old model name from an online example is still available. Prefer the current model list, and check character by character when typing by hand.
Adding it in Mystilink
Open Settings → Models and choose the vendor matching the service. Enter the API key, or, if the interface supports it, the name of an environment variable readable on this machine. With official services, keep the template's default address; overwrite the Base URL only when you clearly need a proxy or a different endpoint.
Test the connection first, then enter or select the model ID, add a display name, and save. After setting the configuration as the primary model, confirm a conversation completes with a message that contains no private records. A successful connection test does not guarantee the model supports the full Agent tool flow.
Vendors and protocols
The following are the corresponding templates and official entry points in Mystilink; they are not a promise about available regions, models, or prices. Key acquisition steps, permissions, and billing rules are defined by the vendor platform.
| Vendor | Official entry | Notes for the Mystilink template |
|---|---|---|
| OpenAI | API platform, authentication reference | Use the OpenAI template; keep the default address and pick a model the account can actually use |
| Anthropic | Console, API docs | The native API uses the Anthropic protocol; do not fill it in as a generic compatible gateway |
| Google Gemini | AI Studio, model docs | Use the Gemini template; the current template does not offer Base URL override |
| DeepSeek | Developer platform, API docs | Use the DeepSeek template and confirm the model and account permissions |
| MiniMax | Developer platform | Use the matching template; do not mix accounts, keys, and addresses across regions |
| Moonshot / Kimi | Developer platform | Use the matching template; the model ID and address should come from the same service region and account |
Anthropic
Choose the Anthropic template when you use the official native service. If a provider also offers a compatible gateway, use the compatible protocol and address it publishes; do not choose the native template just because the model name contains Claude.
Google Gemini
Use the Gemini template to reach the official native API. Gemini models offered by third-party platforms may use a different compatible protocol; choose a custom provider according to the actual gateway.
Address, key, and model must come from the same service
An official key does not necessarily work with a proxy, and a proxy key does not work against the official address. Check all three together: who issued the key, which domain the request goes to, and which models that account can use. After overriding the address, request content goes to whatever service the new address belongs to.
The Base URL should be the API prefix of the service; do not paste a full chat request path or a model detail web page into it. For third-party gateways or local services, see Compatible providers and Local models.
Common problems
| Symptom | Check first |
|---|---|
| 401 / invalid key | Whether the key belongs to this address, whether it expired or was revoked, whether the app can read the environment variable |
| Empty model list | Account permissions and model discovery support; enter a known accurate ID by hand if needed |
| 404 / model not found | The API prefix and the model ID; whether a display name was used as an ID |
| Test succeeds but tasks fail | Tool calling, streaming, and model capabilities, not just the network |
| Quota-related errors | Check API billing with the actual vendor; topping up Newton does not add third-party quota |
More troubleshooting is in Model connection problems. If a key leaks, revoke it on the issuing platform and reconfigure; do not just change the display name or delete a screenshot.