Linux sandbox
Non-sudo user in bubblewrap, LAN blocked by nftables, CPU and RAM limits.
Start free trialFree trial Set what agents may do once, for every machine, and see every attempt to step outside it.
agnte runs every agent in a sandbox. On Linux, agents run as a non-sudo user in bubblewrap with LAN access blocked by nftables and CPU and RAM limits. On Windows, agents run in Docker with a read-only root, all capabilities dropped and an egress proxy that refuses private ranges. A central command policy with 73 denied commands applies to every machine, agents never hold the GitHub token, and every call is journaled.
Non-sudo user in bubblewrap, LAN blocked by nftables, CPU and RAM limits.
Docker with read-only root, all capabilities dropped, internal network and an egress proxy.
73 denied commands, denied paths and web access only for research roles, enforced by a pre-tool hook.
Set or relax sandbox and policy everywhere and see escape attempts in one place.
Owner, operator and read tokens; secrets in the OS vault.
Every agent call and command journaled with its actor.
Source: agnte sandbox self-test and live database, 2026-10-12.
No. The sandbox self-test confirms sudo fails, /etc writes fail, SSH keys are unreadable and LAN connections are refused.
No. Only the control layer writes to GitHub; agents never hold the token.
agnte provides audit-ready logs of every action and decision. It does not hold a compliance certification today.
We do not use tracking or advertising cookies. How we handle your data is explained in our privacy policy.