Introduction

DeepSeek Harness uses a plugin-based architecture. The dsh-llm-agent-bridge plugin uses Anthropic’s official Claude Agent SDK to provide DeepSeek Harness with an optional Claude LLM provider. It allows Claude to be used in any DSH mode, such as Standard, PTC, Creator, Minimal, and more. DeepSeek Harness is responsible for maintaining prompts, tools, the sandbox, and approvals, while Claude handles reasoning and decision-making.

Installation and Prerequisites

Before installing, make sure your system has Node.js 22.13 or later.

Install the plugin:

npx @deepseek-ai/dsh plugin --profile web add github:daveycodez/dsh-llm-agent-bridge
npx @deepseek-ai/dsh web

After installation, you must first authenticate Claude Code as usual. The plugin itself does not read credentials.

Core Capabilities

  • Universal mode support: It can work in any DeepSeek Harness mode (Standard, PTC, Creator, Minimal, etc.).
  • DSH-led control flow: DeepSeek Harness maintains its own prompts, tool definitions, sandbox runtime environment, and approval process.
  • MCP tool bridging: Uses MCP servers to handle tool calls and sends tool requests to DSH for execution.
  • Session telemetry handling: The plugin checks and handles DSH’s session telemetry behavior.

Usage and Configuration

  • Selecting the model: Select Claude from the model dropdown in any DSH mode.
  • Switching models: Supports switching models midway through a session.
  • Configuring thinking summaries: By default, the plugin sets thinkingSummaries: false to preserve the SDK’s default behavior. If customization is required, this setting can be configured on the plugin line.
  • Telemetry control: DeepSeek Harness enables telemetry export by default. To prevent uploading Claude output data, it is recommended to set the environment variable in the shell that starts DSH:
    DSH_TELEMETRY_MODE=DISABLED

Notes and Limitations

  • Text rewrites are non-replayable: If DSH rewrites or edits text after Claude has seen it, the rewritten content will not be included in the Claude session.
  • Session state: The Claude session only knows the turns it has answered. When creating a session or switching models, the adapter prepends previous turns as a <dsh-context> block.
  • Credential security: The plugin does not read credentials.
  • SDK experimental status: The underlying Anthropic Claude Agent SDK methods are explicitly experimental.
  • Guard mechanism: The telemetry check guard runs on every turn rather than at load time.

Conclusion

This plugin integrates Claude into the DeepSeek Harness ecosystem through the official SDK, making it suitable for scenarios where Claude capabilities are needed within a DSH environment. For more details, refer to the GitHub repository or the community catalog.