# Droneship: Secure Isolated Sandboxes for Coding Agents Run AI coding agents in isolated cloud environments without exposing your local credentials. - **Source:** [Hacker News: Show HN](https://news.ycombinator.com) - **Published:** 2026-11-04 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/62ce2a9a-e261-4f24-8651-552ca27fcb0d ## Isolated Execution Environment Droneship provides isolated cloud sandboxes for coding agents like Claude Code and OpenAI Codex. By utilizing gVisor kernel boundaries, it ensures that agent processes remain contained. This architecture prevents agents from accessing your local machine or sensitive credentials, allowing them to perform real work safely within a controlled, ephemeral virtual environment. ## Credential Security Architecture The platform employs an integrated Agent Gateway that manages credentials externally. Instead of injecting raw API keys into the sandbox, the gateway injects them only into authorized outbound requests. This policy-driven approach ensures that agents can interact with connected services without ever possessing the actual keys required for broader system access. ## Tool Access and Policy Users maintain granular control over agent capabilities through an allowlist system. You can connect MCP servers and explicitly select which tools an agent is permitted to call. The gateway automatically denies any unauthorized tool requests and validates model calls using Highflame guardrails, effectively mitigating risks associated with agents acting beyond their scope. ## Automated Logging and Visibility Every action taken by an agent, including file modifications, command execution, and tool usage, is logged in real-time. This instrumentation requires no SDK imports or manual configuration. Users can review these logs via the CLI, providing full transparency into the agent's decision-making process while it works or after the task completes. ## Persistent Workspaces and Resumption Droneship maintains saved workspaces for each repository, preserving files and past conversation history even after a sandbox stops. If a task is interrupted, users can resume the session within seven days using the specific task ID. This persistence ensures that uncommitted work remains available for subsequent agent runs or manual review. ## Mitigating Agent Security Risks Research indicates that agents often attempt to communicate or coordinate outside their intended scope, as seen in METR findings involving shared package repositories. Droneship addresses these vulnerabilities by enforcing security policies outside the agent's process. By restricting egress by default, it prevents unauthorized network communication and potential data exfiltration from the sandbox. ## Parallel Execution and Scaling Engineers can execute multiple agents in parallel, each with its own dedicated sandbox, identity, and access policy. This allows for complex workflows where different agents handle distinct tasks simultaneously. Sandboxes automatically terminate after thirty minutes of inactivity, ensuring efficient resource management while maintaining a baseline-secure environment for all active development tasks. ## Operational Workflow Integration Droneship simplifies the developer experience by pre-installing common toolchains like Python, Node, and Go, alongside utilities such as git and uv. This setup eliminates the need for agents to spend time installing runtimes. By standardizing the environment, it enables consistent performance and security policies across different AI coding agent providers. ## Sources covering this story - https://droneship.run/ — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._