# Agentwares

> Talk to an agent, get a working system. A person describes what hurts in their operation; a team of AI agents understands the problem, writes a Blueprint, and builds a real system (screens, data, rules) on the person's own data.

## How it works
1. Understand: the Guide agent interviews (one question at a time, with options) and reads uploaded spreadsheets.
2. Blueprint: goal, users, process, data, rules and scope on one page. The person approves it. Agents never build outside it.
3. Build: an architect designs data and screens, builders write real Vue screens in parallel, a reviewer checks them. Screens compile in the browser, no build step.
4. Use: the system opens with live data. Changes rebuild only the affected piece. Systems can be shared by public link and remixed.

## For agents: MCP server
- Endpoint: https://sandbox.agentwares.ai/mcp (Streamable HTTP, JSON responses)
- Auth: Authorization: Bearer aw_... (token created at https://sandbox.agentwares.ai/app/account)
- Tools: list_systems, create_system, get_system, send_message, approve_blueprint, read_file, query_data, write_data, set_public
- Flow: create_system -> answer pending_question with send_message -> show the Blueprint to the person -> approve_blueprint -> poll get_system until running is false -> share builder_url.
- Claude Code: claude mcp add --transport http agentwares https://sandbox.agentwares.ai/mcp --header "Authorization: Bearer aw_YOUR_TOKEN"

## Rules agents should know
- Data is never invented by the model: screens read data by reference through a data SDK, with the user's permissions.
- Never approve a Blueprint the person has not seen.
- Public systems are read-only for visitors.

## Links
- Home: https://sandbox.agentwares.ai/
- Pricing: https://sandbox.agentwares.ai/#pricing