Keystone auto-generates dev containers via sandboxed agents

Keystone proves that the most immediate, practical use case for AI coding agents isn't writing new features, but solving the eternal developer headache of environment configuration.

• –By running Claude Code inside a Modal sandbox, it treats the LLM like a disposable sysadmin, completely isolating the host machine from potentially destructive commands.
• –The output strictly adheres to the Development Container standard, making the generated environments instantly usable in VS Code and GitHub Codespaces.
• –It flips the script on project onboarding: instead of developers reading outdated READMEs to set up a project, the repository effectively configures itself.
• –The inclusion of built-in budget controls (defaulting to a $1 limit) directly addresses the valid fear of runaway inference costs during agentic trial-and-error loops.