MangoTart does not sell model tokens or route your requests through a shared MangoTart model account. You bring your own model access, choose it during onboarding, and can change it later from the local Model provider page.
This separation gives you provider choice, but it also means provider availability, limits, billing, and data policies remain between you and that provider.
Your three model-access options
| Option | What you provide | How billing works | Best for |
|---|---|---|---|
| ChatGPT subscription | A supported ChatGPT account through OpenAI's device-code login | Your existing ChatGPT plan; MangoTart adds no model charge | The simplest common setup |
| Provider API key or OAuth | Your own credential for a Hermes-supported provider | The provider charges your account under its own pricing | More provider/model choice and usage control |
| Custom compatible endpoint | Base URL, model ID, and a key if required | Set by whoever operates the endpoint | Self-hosted gateways or compatible services |
The provider list and models shown in the management console come from the installed Hermes version and may change as Hermes and providers change. The current device UI is the source of truth for what your installed version can configure.
Using a ChatGPT subscription
The recommended setup uses OpenAI's device-code flow:
- MangoTart shows a login URL and short code.
- Open the URL on your phone or computer and sign in to ChatGPT.
- Enter the code and approve the connection.
- MangoTart detects the completed login and lets you choose an available model.
Your password is entered on OpenAI's site, not into MangoTart. The resulting login state is stored on the device.
No additional MangoTart model fee
Using an existing ChatGPT subscription does not create a separate MangoTart model charge. Your ChatGPT plan is still a paid third-party service and remains subject to OpenAI's plan availability, limits, acceptable-use rules, and regional access.
Using an API key or another OAuth provider
Choose the provider in the management console, enter the credential it requests, and select a model. Examples include OpenAI, Anthropic, Gemini, DeepSeek, Qwen, MiniMax, and other providers supported by the installed Hermes version.
API use is normally usage-based. The model provider—not MangoTart—sets the price, rate limits, context limits, account requirements, and retention policy. Review those details in the provider's own dashboard before connecting a key.
Provider and tool credentials are stored locally on MangoTart. They are not shown back in plain text in the normal management UI. They can still be included in a manual Hermes backup, and authorized remote support access can expose local credentials. Treat the device and every exported backup as sensitive.
Custom endpoints
If you use an OpenAI-compatible service or your own gateway, choose the custom-provider option and enter:
- the endpoint's base URL;
- the model ID expected by that endpoint; and
- an API key if the endpoint requires one.
Compatibility means the request format is supported; it does not guarantee that every model feature, tool call, or streaming behavior will work identically. Use Test connection and run a simple task before relying on a custom endpoint for scheduled work.
Regional availability
Your model must be reachable from the network where MangoTart runs. In mainland China, OpenAI's ChatGPT device login and inference endpoints are not normally reachable, so the ChatGPT option will fail without a supported network path. MangoTart does not provide a proxy or circumvention service. Choose a reachable domestic or regional provider instead.
The same principle applies elsewhere: provider account eligibility, sanctions, country availability, and network filtering are controlled by the provider and local network—not MangoTart.
Messaging platforms have their own regional availability too. For example, Telegram is convenient where reachable but is not available on every network. See Network and access.
What you may keep paying for
After the one-time MangoTart purchase, possible ongoing costs include:
- Model access — a ChatGPT subscription or usage-based API bill.
- Connected services — Twilio SMS, paid search APIs, storage, data sources, or other services you choose.
- Internet and electricity — normal household or office costs for an always-on device.
- Taxes and import charges — associated with the hardware purchase, where applicable.
There is currently no required MangoTart subscription. Optional services or pricing may change in the future, but they cannot silently become part of an order already governed by the terms shown when you purchased.
Control your spend
MangoTart does not act as your provider's billing meter. To control model spend:
- Set budgets, alerts, or hard limits in the provider's own account where available.
- Start with small, explicit tasks before creating broad schedules.
- Review recurring jobs on the Schedules page.
- Remove unused provider and tool keys.
- Check the provider's actual bill rather than relying on a local estimate.
Scheduled tasks, retries, browser work, and long conversations can all increase usage. A task that runs every hour can cost much more than the same task run once a day.
Switch or repair a model connection
Open Model provider in the local console to see the active provider, change the model, replace a key, or re-authenticate an OAuth account. Use Test connection when available, then restart the agent if the UI asks you to apply the change.
If a model stops working:
- check the provider's service status and your account balance or plan;
- test or replace the credential;
- confirm the selected model still exists and is available to your account;
- re-authenticate expired OAuth access; and
- try a reachable provider if your network blocks the current one.
For a full diagnostic path, see Troubleshooting.
Data sent to the provider
The model needs the prompt and relevant context to perform a task. Depending on the task, that may include conversation history, memory, tool results, webpage content, or text from connected services. Your provider's privacy and retention terms apply to that data.
MangoTart does not proxy these model requests through a shared MangoTart account. See Privacy, security, and permissions for the complete boundary.