Project: https://github.com/chaokunyang/browser-terminal-use
The goal is simple: let an Agent run iterative loops from local code, but execute commands inside a browser-hosted cloud terminal.
With this tool, an Agent can:
- send commands from local CLI - execute inside a bound browser terminal tab - stream output in real time - get the exact remote exit code - use queue / timeout / cancel for stable loops
The key value for me is building a cloud-side verifiable execution environment for local LLM Agents:
- execution happens in a remote/browser terminal context - results are observable and auditable from local workflows - command lifecycle is structured, deterministic, and script-friendly
It is also very useful when using frontier models locally to debug cloud GPU issues (driver/toolchain/env mismatches, flaky jobs, setup regressions) while keeping a tight local Agent loop.
Would love feedback on Agent workflows, reliability across terminal vendors, and ideas for stronger verification primitives.