NanoGPT
Where your scenes get their brains
NanoGPT
NanoGPT is one of RoleCall's partner inference providers — a pay-as-you-go provider that puts hundreds of models from dozens of vendors behind a single API key. Connect it once under Settings → BYOK Providers and every NanoGPT model shows up in the Cast alongside your other providers.
Save 5% with our partner link. New and existing accounts get 5% off when you sign up through nano-gpt.com. The same link sits on the NanoGPT card in the Add Provider screen.
Why NanoGPT
- One key, many vendors. NanoGPT offers OpenAI, Anthropic, Google, DeepSeek, open-weight models, and more — no separate key per vendor.
- Pay-per-use or subscription. Bill per token as you go, or draw from a NanoGPT subscription plan.
- OpenAI-compatible. RoleCall talks to it with the standard chat-completions protocol, so streaming, samplers, and the model picker all just work.
Connecting NanoGPT in RoleCall
- Open Settings → BYOK Providers → Add Provider and pick NanoGPT.
- Paste your API key — get one at nano-gpt.com (5% off through our link).
- Choose a Plan Type:
| Plan Type | What it does |
|---|---|
| Pay-per-use | Bills your NanoGPT balance per token. The default. |
| Subscription | Draws from your NanoGPT subscription quota instead of per-token billing. |
- Optionally pin a Model ID (leave it blank to pick per-scene from the Cast).
- Save, then hit Test Key to validate the key and cache the model list.
The full BYOK walkthrough — encryption modes, multiple configs, custom headers — lives in Providers, BYOK & Encryption and Setting Up BYOK Providers.
Samplers
NanoGPT accepts the standard OpenAI-compatible sampler set:
| Temperature | Top-P | Top-K | Frequency Penalty | Presence Penalty | Max Tokens |
|---|---|---|---|---|---|
| yes | yes | — | yes | yes | yes |
For samplers NanoGPT doesn't expose natively (like min_p), drop them into Custom Body Params and they're merged into every request.
Tool calls
Because NanoGPT serves models from many vendors, native function-calling can be unreliable through it (silent drops, wrong argument types). Leave the in-chat agent's tool-call mode on Auto — it defaults to Prompted, which is more predictable over routers. See Tool-Call Modes.
Good to know
- BYOK usage isn't metered by RoleCall — you pay NanoGPT directly and their limits apply, not your RoleCall tier.
- Test your key after long breaks; provider catalogs and billing change, and Test Key re-primes the model list.