DeepSeek Harness (DSH) Custom LLM Provider Plugin

Introduction¶

The architecture of DeepSeek Harness (DSH) allows functionality to be extended through plugins. When integrating LLM services from specific vendors, several situations often need to be handled: a vendor may require a specific protocol (such as OpenAI Responses or Anthropic Messages), the model list may not be publicly available or may be unstable when fetched dynamically, and compatibility fields may also need to be explicitly declared to accommodate DSH’s internal workflows.

The dsh-custom-provider plugin solves these issues. It provides routing and model directories through static configuration, supports multiple protocols, and reuses DSH’s streaming adapter.

Plugin Overview¶

This plugin is maintained by linziyanleo and licensed under the MIT license. It mainly provides static routing configuration and model directory management. It does not depend on the provider’s /models endpoint and can seamlessly integrate custom LLM providers into DSH’s model selector and standard streaming/tool-calling workflows.

Installation and Activation¶

The plugin must be installed in DSH’s Web profile.

1、Run the following command to install the plugin package:

npx @deepseek-ai/dsh plugin --profile web add @linziyanleo/dsh-custom-provider

2、After installation, verify that the configuration namespace has been loaded successfully by running the following command:

npx @deepseek-ai/dsh --profile web --dump-config

Configuration Methods¶

The plugin can be configured either through the Web interface or via the settings.yaml file.

Web Settings Page¶

1、Open Settings → Custom Providers in DSH Web.
2、Enter the provider ID (must start with a lowercase letter and contain only lowercase letters, digits, and hyphens) and display name.
3、Select a protocol (such as openai-completions, openai-responses, or anthropic-messages).
4、Enter the API base URL and credential reference required by the protocol.
5、Enter the API key. This field is write-only and is stored through DSH’s credential service instead of being written directly to the configuration file.
6、Add models and set parameters such as context window and maximum output.
7、Save the configuration, and the models take effect immediately.

settings.yaml Configuration Example¶

You can also write the llm-custom.providers configuration directly in settings.yaml. The following is an example based on the OpenAI Chat Completions protocol:

llm-custom:
  providers:
    example:
      displayName: Example Provider
      apiKeyEnv: EXAMPLE_API_KEY
      api: openai-completions
      baseURL: https://api.example.com/v1
      compat:
        supportsStore: false
        supportsDeveloperRole: false
        thinkingFormat: deepseek
        supportsReasoningEffort: true
        maxTokensField: max_tokens
        requiresReasoningContentOnAssistantMessages: true
      models:
        - id: example-model
          name: Example Model
          contextWindow: 262144
          maxTokens: 32768
          reasoningEfforts:
            off:
            high: high
            max: max

Protocol and BaseURL¶

The api field determines how the adapter constructs the request path. baseURL must match the expectations of the selected protocol adapter.

Core Features¶

The plugin provides the following capabilities:

• Web management interface: Provides a visual settings page for adding, editing, and deleting custom providers.
• Multi-protocol support: Supports the openai-completions, openai-responses, and anthropic-messages protocols.
• Declarative configuration: Supports declarative configuration via llm-custom.providers in settings.yaml.
• Static model directory: Provides a static model list and does not depend on the provider’s /models endpoint.
• Per-request credential resolution: Dynamically resolves API keys through DSH’s credential service and does not store keys in configuration files.
• Compatibility fields: Supports provider-level defaults and per-model overrides, mainly for the openai-completions protocol.
• Streaming adapter reuse: Reuses DSH’s pi-ai streaming adapter to ensure that text streaming, tool calls, and usage statistics workflows function correctly.
• Full configuration validation: Validates configuration before route registration is replaced, preventing invalid configurations from being partially activated.

Notes¶

• Runtime environment: Requires Node.js ^22.19.0 or >=24.0.0.
• DSH version: Requires the DSH 0.1.0-rc.7 compatible package.
• Profile: Must use the Web profile.
• ID rules: The provider ID in the Web UI must start with a lowercase letter and contain only lowercase letters, digits, and hyphens.
• BaseURL: The baseURL path must match the expectations of the selected protocol adapter (for example, the OpenAI protocol typically requires /v1).
• Compatibility fields: The compat field applies only to the openai-completions protocol and is ignored when switching to other protocols.

Conclusion¶

The plugin adds the ability to configure custom LLM providers to DSH, using static configuration to address the pain points of dynamically fetching model lists and adapting to different protocols. For developers who need to integrate non-standard or vendor-specific protocol providers, it is a practical tool.

• Plugin Directory
• GitHub Repository