jumpserver-dsh

The DeepSeek Harness agent needs to run remote commands, read and write files, and query databases, but dialing asset IPs directly would bypass the permission controls and audit trail that ops teams already enforce. The jumpserver-dsh plugin funnels every outbound flow through JumpServer Core and the KoKo bastion, so command filtering, ACLs, and session auditing keep working unchanged. The agent only ever sees a connection-token and client-url; the real asset IP stays invisible.

Three use cases are covered. On the host side, commands run over SSH exec with full audit, and files move via SFTP. On the database side, sessions go through the KoKo PTY (usql) against MySQL, PostgreSQL, Redis, MongoDB, Oracle, SQL Server, and others; access is read-only by default, and write statements require explicitly enabling the flag. On the asset side, optional host CRUD is available and still gated by Core RBAC. Helper tools such as jms_whoami, session listing, and disconnect complete the loop: list assets, list accounts, connect, execute, disconnect.

Best suited to teams that already run JumpServer in production and want an AI agent to join day-to-day troubleshooting and queries without lowering the security bar. There is no need to open separate SSH or database ports for the agent, nor to maintain an asset inventory on the agent side the permission boundary is defined entirely by the JumpServer Access Key and RBAC policies.