Introduction

DeepSeek Harness (DSH) is a framework that emphasizes “everything is a plugin.” When using Claude Code or Antigravity subscriptions, existing HTTP bridging solutions (such as dsh-subscription-gateway) require separate configuration and running processes, which results in a less smooth experience. dsh-llm-subscription is a native DSH plugin. By registering adapters, it directly integrates Claude, Gemini, and Grok subscription models into DSH’s model selector, without requiring API keys.

Core Features

This plugin mainly solves the problem of “calling large models through subscriptions.” It removes the need to manage API keys manually and provides local free model options.

  1. Subscription model support:
    • Claude: Supports claude, haiku, sonnet, opus, and fable. Each model has a native reasoning effort selector (Low to Max).
    • Gemini: Supports Antigravity paid tiers (Flash Low/Medium/High, Pro Low/High).
    • Grok: Supports grok-4.6 through the official Grok CLI.
  2. Local models: Supports local Ollama models (such as the default qwen3.5:9b), which are free and do not require login.
  3. Response identification: Each response ends with a via <real-model-id> tag, explicitly indicating the actual model used for the current response.

Installation and Enablement

In a DSH profile, run the following command:

dsh plugin --profile web add dsh-llm-subscription

Prerequisites:
* Install and log in to the relevant CLI (claude, agy, or grok), or install and run Ollama.
* Requires DSH version 0.1.1-rc.2 or later.

Typical Usage

After installation, you can directly select the models above in DSH’s model selector to start conversations.

Technical Details and Limitations

The plugin invokes JSON mode output from local CLIs and reassembles the data into the StreamChunk sequence expected by DSH.

  1. No streaming increments: Because the underlying CLI does not support streaming output, responses are returned as a single block.
  2. No exposed thinking content: The reasoning-chain text cannot be retrieved; only token counts are visible. The reasoning effort selector can still control the amount of computation.
  3. Gemini tool isolation: Unlike Claude, it cannot isolate tools using --strict-mcp-config. Gemini is aware of agy’s built-in tool set.
  4. No image input support.
  5. No credential handling: It does not store or transmit keys, and relies entirely on locally logged-in CLI sessions.

Security and Compliance Risks

Important note: This plugin carries account-ban risk.

Anthropic’s official terms explicitly prohibit third-party developers from routing requests on behalf of users, especially using credentials from Free, Pro, or Max plans. DSH is not an unmodified Claude Code binary, but a standalone framework that uses the plugin to invoke claude -p as its backend. This pattern violates current compliance rules. Although this currently mainly targets Anthropic, Google’s penalties for Antigravity/Gemini violations are already very strict.

Grok is the only provider without such compliance risk, because xAI officially supports third-party subscription OAuth integration.

Summary

dsh-llm-subscription provides a native way to integrate Claude, Gemini, and Grok subscriptions into DSH, and adds local Ollama support. It removes the burden of managing API keys, but note the lack of streaming output, hidden reasoning content, and potential compliance-related account-banning risk.