deepseek-harness-ssh

When AI agents need access to isolated laboratory GPU servers, traditional SSH terminals can be dangerously permissive, exposing arbitrary commands and file paths. deepseek-harness-ssh addresses this critical security gap by providing a safety-gated SSH development toolkit for DeepSeek Harness and standard MCP clients. By implementing a strict "deny-by-default" policy, the plugin ensures that agents can only interact with pre-configured fixed host aliases, protected filesystem paths, and command allowlists, maintaining operational safety while preserving server isolation.

The plugin offers robust core capabilities, including fixed SHA256 host-key fingerprint verification, alias-based remote directory root constraints, atomic file writes with hash protection, and command execution timeouts. It also features an HTTPS artifact bridge specifically designed for offline servers, enabling secure artifact staging and validation without exposing the infrastructure to the public internet. As of version 0.3.0, the DeepSeek Harness plugin, stdio MCP server, and command policy are fully implemented, backed by a test suite of 25 tests to guarantee isolation from real infrastructure.

This solution is ideal for developers, researchers, and DevOps professionals who need to run AI models in a secure sandbox. If you require access to laboratory computing resources but must strictly prevent unauthorized modifications or access, this tool provides granular control, ensuring your server remains completely isolated from the public web.