Agentsh

agentsh, short for agent shell, is an execution layer security gateway for AI agents. It sits underneath an agent or harness, intercepts file, network and process activity at the system call level, and applies a deterministic policy to every operation regardless of what the prompt, tool output or user says. The premise is that models are probabilistic while execution should not be. It is built for teams running coding agents and autonomous workloads who need guarantees that do not depend on prompt compliance. Each action can be allowed, denied, paused for human or CI approval, redirected to a safe alternative, audited with detailed logging or turned into a soft delete. Every file, network and process operation is captured, including subprocess trees spawned by package installs, npm scripts and makefiles, and structured events can be routed to any logging system. This addresses prompt injection, jailbreaks, reasoning errors and secret exposure through environment variables or mounted credentials. It installs on macOS and Linux and works with agents such as Claude, GPT and Cursor, and a secure sandbox SDK for npm and pip adds agentsh to hosted sandboxes including E2B, Daytona, Vercel, Modal, Cloudflare, Deno, Runloop and Freestyle. agentsh is open source under the Apache 2.0 license and free to install from GitHub. No paid plans or prices are published; the site offers an Enterprise contact route for organizations that want to discuss commercial terms.