AgentSandbox — Self-hosted sandbox-as-a-service for coding agents

Run coding agents in fully isolated containers without touching your machine

Problem

Many developers run coding agents (like pi coding agent, Claude Code) on their own servers but worry about destructive commands or accidental local file modifications. Existing Docker sandbox setups require manual stitching, are tedious to configure, and lack unified management.

Solution

A lightweight self-hosted console that spins up isolated microVMs/containers per agent task with one click, auto-injects repository code, restricts network and filesystem access, and destroys the sandbox after completion. Supports task submission via web UI or CLI with real-time agent execution logs.

Users

Individual developers and small teams who heavily use coding agents, own self-hosted servers, and care about security isolation

MVP

Integrate at least one coding agent (e.g., pi or Claude Code); provide one-click sandbox creation via CLI and web UI; auto-clone specified repo inside sandbox; restrict sandbox to working directory only; auto-destroy and reclaim resources after task

Revenue

Open-core with paid team-hosted version or enterprise support subscription, priced by concurrent sandbox count

Why now

Coding agents are moving from toys to daily tools, yet security isolation remains manual scripting; underlying tech like Docker and Firecracker is mature, letting indie developers quickly wrap an easy experience