Gadget shows how it built better AI support drafts
Gadget shared an internal system that lets AI investigate support tickets and write first drafts. The important idea is giving the AI real product context, not just public docs.
Many AI support bots can answer questions that are already covered in docs, but they struggle with real app problems like errors, logs, or high usage. Gadget’s setup creates an E2B sandbox for each ticket, then uses the Claude Agent SDK to inspect code and draft a reply.
The sandbox includes Gadget’s monorepo, the user’s app code, analytics, and a custom MCP server that connects the AI to tools a human support engineer would use. Gadget says this is meant to speed up human support, not replace it. One practical point is scaling: if 1,000 tickets arrive, the system can create 1,000 separate sandboxes so work can happen in parallel.
Key points
- Useful AI support needs real product context, not only help docs.
- Each ticket gets its own E2B sandbox so the agent works in an isolated space.
- The Claude Agent SDK drafts answers after reading the question and relevant code.
- The MCP server lets the AI use internal tools and app data like a support engineer.
- A human still reviews and edits the final response before sending it.
Quick term guide
- context
- The information an AI uses to understand your request, such as files, notes, and past messages.
- E2B sandbox
- A separate, temporary workspace where an AI agent can safely run and inspect things.
- sandbox
- A separate space where code can run without affecting other work.
- Claude Agent SDK
- A developer tool for building AI agents that use Claude to work through tasks.
- monorepo
- One code storage place that holds code for several apps or services.
- analytics
- Stats that show things like visits, clicks, and user activity.
- MCP server
- A server that helps AI tools connect to outside services in a standard way.
- server
- A computer that stores files and shares them with other devices in your home.