DeepSeek Harness (dsh) Setup Guide

DeepSeek Harness is DeepSeek's open-source agent runtime framework (command dsh), built around the core idea of "everything is a plugin": models, tools, sessions, and other capabilities are all provided by plugins. It supports connecting to any gateway through a "custom model API" — NextHub's GPT-Azure group documentation already lists DSH as a compatible third-party tool.

Connection Steps

1. Install and Launch

You need Node.js installed (the current official LTS release is recommended). Start the Web UI directly with npx, or install it globally first and then start it:

npx @deepseek-ai/dsh web

npm install -g @deepseek-ai/dsh
dsh web

Once launched, your browser automatically opens the local management interface.

2. Add a Custom Provider (Claude Channel)

Go to Settings → Models, click Add Provider, switch the card to Custom model API, and fill it in according to the table below:

Field	Value	Notes
Provider ID	nexthub	Lowercase; cannot be renamed after creation
Base URL	https://api.nexthub.cc	Bare domain for the Anthropic protocol, without /v1
API protocol	Anthropic Messages	Corresponds to the Anthropic Messages protocol endpoint
Credentials (API Key)	your NextHub ApiKey	A Claude group
Models	claude-opus-5-5, etc.	At least one; click "Fetch available models" to pull them automatically

After saving, select the model you just added in the conversation's model list to use it.

3. Connect the GPT Channel (Optional)

A provider speaks only one protocol, so if you need GPT, create another one: set the Provider ID to nexthub-gpt, choose OpenAI Chat Completions (or OpenAI Responses) for the API protocol, fill in https://api.nexthub.cc/v1 (with /v1) for the Base URL, use a GPT-Azure / GPT-Pro group key as the API Key, and set the model to gpt-5.6, gpt-6-astra, and so on.

4. Verify

Select a NextHub model in the conversation and send a message. If you get a normal reply, you are connected.

Configuring via File (Optional)

The configuration above is ultimately written to $DSH_HOME/profiles/web/cordis.patch.yml (DSH_HOME defaults to ~/.dsh); you can also edit this file directly:

- id: llm-pi-ai
  config:
    providers:
      nexthub:
        apiKeyEnv: NEXTHUB_API_KEY
        api: anthropic-messages
        baseURL: https://api.nexthub.cc
        models:
          - id: claude-opus-5-5
          - id: claude-fable-5

apiKeyEnv takes the name of an environment variable (the key itself is read from that environment variable); keys saved through the UI are instead stored in $DSH_HOME/.credentials.yaml. Changes take effect on the next request without a restart.

Vision models must declare input types

Manually entered models are treated as text-only by default. To send images to a model, check Image under Input types in the model's Model options (corresponding to the config input: [text, image]); otherwise the image is rejected before it is sent.