99-second demo: controlled terminal + browser agent execution via MCP with approval boundaries
ImaginaryMachine9110 · reddit · 2026-09-29
An indie project demos letting an AI agent do real work across terminal and browser via an MCP layer without blanket machine access. A 99-second video shows a cross-surface workflow where the MCP layer separates capability from permission: policy checks before sensitive actions, controlled local operations instead of unrestricted access, and observable/auditable execution. The author is soliciting community input on the minimum approval boundaries before letting agents operate real machines.
More from coding & agent
- IGSD Uses Environment-Verified Hindsight Self-Distillation to Train Search Agents — _reachsumit · 2026-09-29
- Prompting a Coding Agent to Build a Music Video Overnight — neoneye2 · 2026-09-29
- AI agents: you want sociopathic geniuses writing your code, not reading your email — tobowers · 2026-09-29
- StreamChat SwiftUI 5.12.0 ships iPhone Duo support with coding-agent workflow — amos_gyamfi · 2026-09-29
- Codex now edits code via temp files, hiding which files were changed — xeophon · 2026-09-29
- Simon Willison publishes annotated notes for his "2026 in LLMs (so far)" keynote — AccBalanced · 2026-09-29