Providers
A SharpMind Compute host is the combined server image: the Studio web UI on 5117 and the OpenAI-style API on 8080, sharing /data/SharpMind. The provider only supplies the machine, Docker, ports and disk.
Compare hosts
| Host | Setup | GPU | Billing | Best for |
|---|---|---|---|---|
| Local Docker | Studio starts the container | Optional local NVIDIA GPU | Free (your power) | Privacy, small models, CPU pretraining |
| Runpod | Studio deploys the pod via API | Yes | Per minute; balance required | Short GPU jobs without a server to keep |
| Vultr | Studio deploys the instance via API, then sets up over SSH | Yes, A16 class | Hourly; also while stopped | A host you keep for a day of experiments |
| Vast.ai | Create a GPU instance; paste its URL and token | Marketplace GPUs | Per minute | Short GPU jobs on marketplace capacity |
| MaxCloudOn | Create a GPU instance; paste its URL and token | Cloud GPUs | Provider plan; code SharpMindStudio for 20% off | A managed GPU host outside your machine |
| Novita | Create a GPU instance; paste its URL and token | Cloud GPUs | Provider plan | GPU hosting with an optional inference API |
| Custom / bare Linux | Your Docker + optional SSH | Whatever the box has | Your provider | An existing server, workstation or lab machine |
Sign-up links support the project and do not change the price. MaxCloudOn code SharpMindStudio takes 20% off.
Remote hosts and routers
Remote hosts rent you a machine; routers provide hosted model endpoints. Vast.ai, MaxCloudOn and Novita have manual remote-host presets under Settings > Providers > Compute; paste the instance URL and token after creating the GPU. MaxCloudOn code SharpMindStudio takes 20% off. OpenRouter, EuroRouter (beta) and OpenCode Go are integrated in Studio; each keeps its own key, base URL and model list.
| Provider | Kind | Status | Link |
|---|---|---|---|
| Vast.ai | Remote host | Manual host preset | Sign up |
| MaxCloudOn | Remote host | Manual host preset · code SharpMindStudio (20% off) | Sign up |
| Novita | Remote host / inference | Manual host preset | Sign up |
| OpenRouter | Router | Integrated | Guide |
| EuroRouter | Router | Integrated (beta) | Sign up |
| OpenCode Go | Router | Integrated | Sign up |
EuroRouter keeps a profile beside OpenRouter: https://api.eurouter.ai/api/v1, an eur_... key and its own model list. Test key validates the key without spending credit and reports the remaining balance. The beta badge stays until a live completion has been verified.
OpenCode Go uses https://opencode.ai/inference/go/openai/v1 and a Console API key. Studio sends x-opencode-session per chat for routing and prompt cache, and Test key checks the model catalogue because Go has no balance endpoint. See Other providers.
The shared contract
Every host follows the same rules, which is why the workflow does not change:
| Item | Value |
|---|---|
| Studio UI | 5117 (and 5118 for the LAN listener) |
| Compute API | 8080 |
| Data root | /data/SharpMind mounted from the host |
| Models | /data/SharpMind/Models |
| Plugins | /app/plugins (the GPU image carries SharpMind.GPU) |
| Auth | SHARPMIND_TOKEN on every API call; /health is open |
| Default model | One active model per process; activate another from Studio or the API |
Which one?
- Only trying the app: run Studio locally and use a packaged model. No host needed.
- Want a GPU for a fine-tune this week: Vultr. Keep the instance for the session, destroy it when done.
- Want a GPU for a few minutes and already have credit: Runpod, deployed from the card.
- Have a Linux box with Docker: custom host; point the card at it and press Test.
- Need full privacy: Local Docker or the local Studio process.
What SSH setup does
For Vultr and custom Linux hosts, Studio can log in over SSH and prepare the machine. It is optional: a host that already runs the image only needs a Base URL, token and Test.
For an API deploy, Studio only uses SSH keys that already exist on the Vultr account; it never uploads one. Create or copy the key in Edit, add it once in the Vultr console, then select it in the Vultr panel. See Add an SSH key.
The helper performs these steps, one at a time, with streamed output:
- Probe the host: OS, architecture, Docker version, NVIDIA GPU and free disk.
- Install Docker from get.docker.com when it is missing.
- Install the NVIDIA container toolkit when the GPU host checkbox is on and it is missing.
- Pull the server image. The image name must include a registry prefix, for example integral2u/sharpmind-server:gpu.
- Replace the container, mount /data/SharpMind, publish 5117 and 8080, and pass the saved token.
- Inspect and test the container, then call /health from Studio.
It deliberately does not install the NVIDIA driver: that needs a reboot and can break the session. If nvidia-smi is missing, setup stops and says so. Use a provider GPU image that includes the driver, or install it once and run setup again.
Tokens and security
- The host token is generated in Studio and stored encrypted with the local key box.
- SSH passwords and key passphrases use the same encryption; private keys stay on this PC.
- The SSH host key is trusted on first use. If that matters for your environment, pin it manually.
- Rotate the token before opening a host to the internet; a literal value such as PASTE_YOUR_TOKEN is not a secret.
- Restrict the provider firewall to your own IP where possible. SSH 22, API 8080, Studio 5117.