Introduction

Integrating an LLM Provider into DeepSeek Harness (DSH) involves more than just setting the baseURL; common tasks include handling model lists, protocol differences, context windows, output limits, reasoning controls, and credential refreshes. DSH’s philosophy is “everything is a plugin”; the community catalog is an independent site with no official affiliation to DeepSeek / Huaxuan, so it should not be misconstrued as an official app store.

Below, I introduce OpenSaozi/dsh-xai-catalog. It is a model catalog plugin for DeepSeek Harness targeting the xAI and Grok Coding Plan LLM Provider, aiming to integrate available Grok models into DSH and provide verified configuration presets for Grok 4.6.

What is this

OpenSaozi/dsh-xai-catalog is provided via the OpenSaozi GitHub repository under the MIT license. It is an independent integration with no affiliation or endorsement from xAI, DeepSeek, or pi-ai maintainers. Grok and related model names are trademarks of their respective owners.

The problem this plugin solves is specific: to enable users holding an xAI API key or xAI subscription to use Grok 4.6’s coding and reasoning capabilities within DeepSeek Harness.

Core Capabilities

Based on verified information, this plugin primarily provides the following capabilities:

  1. Query the xAI /models endpoint to dynamically display Grok models available to the current account.
  2. Provide verified configuration presets for Grok 4.6, including context windows, output limits, and reasoning controls.
  3. Register its own xai adapter: grok-4.3 and grok-build-0.1 use Chat Completions; grok-4.5 and grok-4.6 use Responses.
  4. Read JSON credentials via XAI_OAUTH_CREDENTIAL, reuse pi-ai’s built-in xAI OAuth handler to refresh expired tokens, and write back the rotated credentials via Harness credential service.
  5. Keep OAuth JSON outside of Cordis configuration and logs; only the handler-derived access token is passed as a request-level auth override to the exported PiAiAdapter.
  6. Does not inject additional prompts, headers, token overhead, or transport alterations.

It is necessary to distinguish between “model availability” and “availability of protocol details.” The xAI /models endpoint can prove that a model ID is available to the current account, but it does not directly prove protocol, context capacity, modalities, or reasoning controls. New model IDs will only be exposed after maintaining the corresponding descriptor.

Installation and Usage

During installation, you need to specify the DSH profile and the commit revision to install. First, replace <profile> and <commit-sha> in the command, then execute the following command:

dsh plugin --profile <profile> add github:OpenSaozi/dsh-xai-catalog#<commit-sha>

This step installs the reviewed revision into the specified DeepSeek Harness profile.

The repository contains a dsh.bundle patch and prebuilt runtime artifacts, so the installation process does not require local building.

The installed bundle provides the following Cordis configuration:

- id: model-catalog-xai
  name: '@deepseek-ai/dsh-model-catalog-xai'
  config:
    oauthCredentialEnv: XAI_OAUTH_CREDENTIAL
    baseURL: https://api.x.ai/v1
    displayName: xAI

This configuration explains the plugin’s entry ID, credential environment variable, baseURL, and display name.

Typical Usage

This plugin is suitable for users who have already completed xAI credential preparation. The usage flow is roughly as follows:

  1. Prepare an environment where DSH profiles can be written.
  2. Install the specified commit using the command above.
  3. Ensure XAI_OAUTH_CREDENTIAL points to a stored xAI JSON credential reference.
  4. View Grok models exposed by the plugin in DeepSeek Harness’s model selection.
  5. Select Grok 4.6 for coding or reasoning-related tasks.

Credential refresh is triggered when the plugin is mounted and when XAI_OAUTH_CREDENTIAL changes. If refresh fails or authentication is not passed, the plugin will retain the maintained catalog and record the error, replacing the model list with empty data.

Note that this plugin does not provide a login UI. It reads credentials already stored in XAI_OAUTH_CREDENTIAL and refreshes them itself; re-login still requires the deployment side to write the pi-ai xAI handler result back to that reference.

Applicable Scenarios and Notes

The suitable user persona is clear: users who already have an xAI API key or xAI subscription and wish to use Grok 4.6’s coding and reasoning capabilities in DeepSeek Harness.

Before use, it is recommended to pay attention to the following points:

  1. This is an independent integration and does not represent an official product, official catalog, or official endorsement of xAI, DeepSeek, or pi-ai.
  2. The plugin runs with the current dsh process permissions; check the source code, dependency scope, and MIT license before installing.
  3. New model IDs will not automatically receive full protocol configuration just because xAI /models returns them; maintaining the corresponding descriptor is required.
  4. The credential security model depends on the deployment method of XAI_OAUTH_CREDENTIAL; the plugin itself is not responsible for the login UI.
  5. Names such as Grok, grok-4.3, grok-4.5, grok-4.6, and grok-build-0.1 should retain their original casing to avoid confusion with other models.

Conclusion

The value of OpenSaozi/dsh-xai-catalog lies in centralizing the xAI/Grok model catalog, protocol adaptation, Grok 4.6 configuration presets, and credential refresh into a single DSH plugin, reducing the configuration complexity of integrating Grok into DeepSeek Harness.

Current verified information does not provide a catalog page URL, so the catalog page address is not listed here. The repository address is as follows:

https://github.com/OpenSaozi/dsh-xai-catalog