DeepSeek Harness (DSH) aims to extend capabilities through a plugin system. In operations scenarios, agents or developers often need to directly operate servers during a conversation instead of only handling text. The dsh-withSsh plugin integrates AI collaboration, real SSH terminal operation records, and operation traces on a single page, providing a visual console.

This is a client-side plugin maintained by lhwu1, licensed under the MIT License. It is not a standalone SSH client, but exists as a DeepSeek Harness plugin, used to connect to servers within the conversation and allow humans and AI to share a terminal for operational tasks and review.

Core Features

The plugin primarily provides the following capabilities:

  • Server status monitoring: provides entry points to view key metrics such as CPU, load, memory, and swap.
  • Log troubleshooting and service checks: supports entering commands directly in the terminal and letting the AI execute commands and return results.
  • Compute and process checks: used to confirm GPU or NPU usage.
  • Directory browsing: provides a read-only file tree view, displaying directories and file metadata.
  • Human-AI collaboration: humans and AI can execute commands in the same conversation, and operation sources are clearly distinguished in the terminal.

Installation and Enablement

Installing the plugin requires DeepSeek Harness to be installed and dsh to be usable from the command line. Add the plugin to the Harness web profile:

dsh plugin --profile web add github:lhwu1/dsh-withSsh

After installation, verify that the configuration has been loaded. Ensure the output contains dsh-withssh and dsh-withssh-web. Then start the Web frontend:

dsh --profile web --port 3080

Open the corresponding URL in a browser and select the SSH Console tab in the conversation to use it.

Typical Usage

  1. Establish a connection: In the conversation’s SSH console, enter the host, port, account, and password to establish the connection. The plugin returns the SHA-256 host key fingerprint, which must be manually verified and confirmed by the user.
  2. Manual operations: After the connection succeeds, directly enter commands in the terminal at the top right (such as top and systemctl) to check service status or logs.
  3. AI collaboration: In the conversation, tell the AI which service needs to be checked. The AI executes commands using ssh_exec. AI commands in the terminal are displayed in yellow; click ? to view the Chinese description.
  4. Result interaction: The terminal displays remote echo, and the AI conversation also receives standard output, error output, and the exit code. The user decides the next step based on real logs.

Applicable Scenarios and Notes

  • Target audience: operations personnel, SREs, platform engineers, and developers who need to collaborate with AI for server inspection and troubleshooting.
  • Security and limitations:
    • The password is only temporarily passed through the connection request and the plugin process memory; it is not written to a persistent file.
    • The host key fingerprint must be manually confirmed during the first connection.
    • Remote output is limited to 256 KiB, and ANSI/OSC control sequences are removed.
    • The file browser is read-only and does not support uploading or deleting files.
    • High-risk operations (such as recursive deletion) require Harness approval.
    • The plugin relies on the Harness Web frontend’s same-origin access model; restarting Harness clears the SSH connections in memory.

By integrating AI conversation with a real terminal, this plugin addresses the need to perform server operations and review them in a single interface. For more details, see the plugin directory or source repository.