The Agent Skills Directory

[COMMAND_EXECUTION]: The SDK provides methods to execute shell commands within the managed cloud sandbox for tasks such as running tests or build scripts.
Evidence: box.exec.command("npm test") in SKILL.md.
[REMOTE_CODE_EXECUTION]: Facilitates the remote execution of code snippets in Python, JavaScript, and TypeScript within isolated containers.
Evidence: box.exec.code(code="print(1 + 1)", lang="python") in SKILL.md.
[EXTERNAL_DOWNLOADS]: Supports downloading agent skills from GitHub repositories and integrating Model Context Protocol (MCP) servers from package registries. The examples use trusted vendor repositories and official packages.
Evidence: skills=["upstash/qstash-js"] and mcp_servers=[{"package": "@modelcontextprotocol/server-filesystem"}] in SKILL.md.
[DATA_EXFILTRATION]: Features capabilities to expose sandbox ports via public URLs and send run results to webhooks. These are intended developer features for service integration and previewing.
Evidence: box.get_public_url(3000) and webhook={"url": "https://example.com/hook"} in SKILL.md.
[PROMPT_INJECTION]: The skill exposes a surface for indirect prompt injection as it processes natural language prompts that can trigger tool usage (shell, filesystem) within the sandbox.
Ingestion points: Prompts provided to box.agent.run and box.agent.stream, as well as repository content from box.git.clone.
Boundary markers: None explicitly defined in the SDK usage examples; agent security relies on the chosen harness (e.g., Claude Code).
Capability inventory: Comprehensive access including shell command execution, filesystem writes, and network-enabled webhooks.
Sanitization: Not demonstrated in the provided SDK usage examples.

upstash-box-py