Tweet
my mind is blown at how cool this is, there's no world CLIs are better than this: - you can bring any MCP, API, tool and it becomes callable from typescript - the only thing that is in context for the agent is the tool `run_code` - for APIs, all GETs are allowed by default - other operations require approval - it has a discover function it can call to get LSP information about input / output types - you can use the input type information to dynamically generate approval rules that match the shape of the function - it runs in a sandbox, so you can allowlist hosts in this demo it's deleting all domains from a project of mine except one, and it literally can't go off the rails due to the approval model
This has a linked resource worth reading, a tool worth trying, or an idea worth prototyping
Quick Insight
Rhys is showing off an AI agent system that can execute TypeScript code and call APIs with built-in safety guardrails - requiring approval for destructive operations while auto-allowing safe ones like GETs. The sandboxing and approval model solve the "AI agent going rogue" problem that makes most developers hesitant to give agents real system access.
Actionable Takeaway
Try building a safe AI agent for one of his side projects - maybe automating print-on-demand product management or web agency client tasks where the agent can read data freely but needs approval before making changes or spending money.
Related to Your Work
This could be huge for the fintech platform's webhook debugging and analytics work - an AI agent that can safely explore customer data, analyze patterns, and suggest fixes while requiring approval before touching live payment flows or customer accounts.
Source Worth Reading
No links provided, but this sounds like a specific tool/framework demo. The thread might contain implementation details or tool names worth investigating for the actual system being shown.