Vercel ships Run SDK: sandboxed eval for agent code with human-approval pause/resume
cramforce · x · 2026-08-26
Vercel released Run SDK, a package for executing untrusted JavaScript/TypeScript generated by agents:
- Code runs in a fresh QuickJS sandbox inside a worker thread, with no direct route to Node.js, the network, or app secrets.
- Apps expose only approved operations via hostFunctions, which become callable globals inside the sandbox.
- Execution can pause at authentication or human-in-the-loop approval boundaries, then resume without repeating completed work.
- Install with pnpm add run; runs anywhere Node.js runs, already shipped in AI SDK as the new primitive for code mode.
More from coding & agent
- Peter Yang open-sources /fuck-cancer, an AI skill helping patients advocate for themselves — petergyang · 2026-08-26
- User Rants: Claude Outperforms Codex in Coding Tasks — aloncarmel · 2026-08-26
- OpenAI desktop app adds WebMCP support, enabling Codex to deploy apps to Sites — OpenAIDevs · 2026-08-26
- OpenAI Introduces WebMCP for Direct Agent-Web Tool Integration — OpenAIDevs · 2026-08-26
- devOS: MCP Server for Collaborative Memory in Coding Agents — black_phoenix9 · 2026-08-26
- Agentmuxer launches as an 'OpenRouter for Agent Capabilities', pay-per-use with no keys — ycombinator · 2026-08-26