What problem does it solve? Agents that run bash, read, write, and edit tools on the host machine lose their work on restart and clutter the local environment. This Skill binds those tools to a remote sandbox per conversation, so work survives restarts, is browsable from a file UI, and is reclaimed on a timer. ## Core Features & Use Cases - Harness Bindings: Provides hands bindings for Claude Agent SDK, OpenCode, and a generic MCP binding for any other harness, all sharing one core behavior. - Platform Setup Guidance: Walks through the prerequisites that must exist first — an env with an E2B-compatible template, a pool sized to concurrent conversations, and a tenant key for the right identity. - Identity & Credential Model: Explains per-person versus service-owned sandbox identity, and how decoy credentials with egress-sidecar substitution keep real secrets out of the sandbox. - Use Case: You run a Claude Agent SDK-based product and want each user conversation to execute shell commands in its own durable sandbox instead of on your server. This Skill tells you which binding to wire, how to size the pool, and which identity model to choose. ## Quick Start Ask the agent to connect your Claude Agent SDK harness's bash and file tools to an AgentBox sandbox for the current conversation.
