I built Capsule Bash, a sandboxed Bash made for agents without setup

Existing Bash is quite annoying to set up securely. We either need Docker or use some cloud-based solutions. On top of that, standard Bash doesn't always return feedback after a command to help enrich the context history. I've been rebuilding a more suitable Bash for agents. It doesn't include all Bash commands and shell features, only the ones that suit agents' needs. **Sandboxing** Every execution is sandboxed in an isolated WebAssembly environment (via WASI 0.2 specifications). This becomes essential for commands like `python3` or `node`, available in Capsule Bash, that might execute untrusted code. I wanted to make sure you can download anything and execute it without risking the host system. **Enriched returns** The idea is to give structured feedback on each command executed, showing what was created, modified, or deleted. It also helps keep a trace in conversations so the LLM can understand context when resuming. If an agent runs `rm -rf`, it will know directly what actions were performed; no need to run `ls` afterwards. \-- The usage is quite simple: import { Bash } from '@capsule-run/bash'; import { WasmRuntime } from '@capsule-run/bash-wasm'; const bash = new Bash({ runtime: new WasmRuntime() }); const result = await bash.run('echo "Hello from sandboxed bash"'); I built it in TypeScript with the \`WasmRuntime\` in Rust. It also has an MCP version if that fits better. More details on how it works are on the main repository: [https://github.com/capsulerun/bash](https://github.com/capsulerun/bash) I'd love to know what you think about this approach.