Sentient

What makes it different

After months of working rigorously with Claude Code, the maker of Sentient caught himself blindly approving permission requests just to keep his projects moving — and realized coding agents run hundreds of commands underneath the covers that he never saw. Agents have deleted databases and production environments by accident, and the news made the risk concrete. So he built something to gain visibility into what is actually happening, and potentially block things before they occur.

Sentient keeps an audit trail of what an agent actually did, from the kernel up: every file change, shell command and network connection is captured, so you can reconstruct exactly what happened and when — including the commands that never show up as edits in the chat. It surfaces the risky operations for review. From the maker's own coding agent history: 5.4k shell commands across 19 sessions, a median length of 132 characters, a longest command over 10k characters, and around 678 commands that modified files.

Sentient is open source and runs on Linux, with a download and docs on nosaka.ai and the code on GitHub. It is built for developers who let agents work in their repos and want ground truth about what changed, rather than trusting the chat transcript.