DeepSeek Harness (dsh) Gateway Setup Guide (2026)

DeepSeek Harness (dsh) is DeepSeek's official open-source agent runtime, released on 13 August 2026 alongside V4-Pro and passed 100,000 GitHub stars within two days. Its design philosophy fits in one sentence: everything is a plugin — including the agent loop itself. That makes dsh less a fixed coding assistant and more a modular runtime for building coding agents.

Why put a gateway behind dsh

dsh is model-agnostic: a custom provider needs just a base URL, a protocol and model IDs. Connecting to every vendor directly means juggling keys, invoices and network paths; through one HopBase gateway the same dsh install runs DeepSeek, GLM, Kimi, Grok, GPT and Claude side by side, with quotas, call auditing and cost attribution in one console.

Three-minute setup

Add a provider to $DSH_HOME/settings.yaml:

providers:
  hopbase:
    api: openai-completions
    baseURL: https://api.hop-base.com/v1
    apiKeyEnv: HOPBASE_API_KEY
    models:
      - id: deepseek-v4-flash-202605
      - id: glm-5.3
      - id: grok-4.6

Export the key before launching: export HOPBASE_API_KEY=sk-your-key. Prefer a UI? In dsh Web UI go to Settings → Models → Add a custom provider and fill in the same three values; take model IDs from GET /v1/models.

Protocol cheat sheet

Model family	`api`	`baseURL`
DeepSeek / GLM / Kimi / Grok / GPT chat	`openai-completions`	`https://api.hop-base.com/v1`
grok-4.20-multi-agent (Responses-only)	`openai-responses`	`https://api.hop-base.com/v1`
Claude models	`anthropic-messages`	`https://api.hop-base.com` (no /v1)

What we verified

The two things agents depend on are verified end to end against the HopBase production gateway: function calling (dsh-shaped tool requests return proper tool_calls) and streaming usage reporting (the final chunk carries full usage; billing reconciles line by line with the upstream).

Quick troubleshooting

Error	Cause	Fix
404 "platform does not support this API path"	Claude model on the OpenAI protocol	Use `anthropic-messages` with the base URL without /v1
model not found / 503	Model not in your plan key's group	Use the exact IDs from `GET /v1/models`
401 invalid api key	Truncated key or stray whitespace	Re-copy the key from the console

Full parameters and more clients in the DeepSeek Harness integration guide; sign up to create a key and start.