Give each of your users a persistent agent computer, with fast starts and your own models.
Create one template for your product and one Sanbox Computer per user, keyed by your user ID. Stop it when they leave, resume it when they return, and use warm pools for fast first starts.
Prepare a Computer once, snapshot it, and keep a warm pool of ready copies, so a new Computer can start in under a second. Snapshot templates let you promote or roll back.
Add the endpoint as a Custom OpenAI Proxy with sanbox model-providers create, then build templates on it. Any OpenAI-compatible URL works, such as an Azure OpenAI EU Data Zone deployment.
Connect your Supabase project to Sanbox once, then create a Hermes Computer per user with their Supabase Auth ID. The user approves access, and every read runs as them under your RLS policies.
The Sanbox MCP server lets Claude, ChatGPT, Cursor and other MCP clients start and use Sanbox Computers as tools. It's enabled on request; today, agents with a shell use the Sanbox CLI.