Foxwire Logo
DeepSeek Harness × Sozdai AI

Run DeepSeek Harness on Sozdai AI

DeepSeek's official agent framework with a backend you can pay for: same V4 models, plus Claude and Kimi. Three steps, all in the web UI.

3-step setup
No DeepSeek account needed No config files ~3 minutes
DeepSeek Harness Models settings page

Screenshot: official DeepSeek Harness docs (MIT)

Official harness, payable backend

DeepSeek's own agent framework, powered by an API you can actually pay for.

100% GUI setup

No config files: add a provider in Settings, models auto-load via Fetch.

One key, more than DeepSeek

Same key also unlocks Claude, Kimi, Qwen and GPT in the model picker.

Pay per token

One balance, no subscriptions, prices from 24 ₽ per 1M tokens.

Official DeepSeek API vs Sozdai

Official APIVia Sozdai
Access from RussiaAPI returns 403 for Russian IPsDirect access
PaymentNo Russian cards (CNY/foreign cards)Russian cards, rubles
ModelsDeepSeek onlyDeepSeek V4 + Claude, Kimi, Qwen, GPT
AccountDeepSeek account requiredOnly a Sozdai account

Set up in 3 steps

Everything happens in the web UI — no files to edit.

  1. 1

    Create a Sozdai API key

    Sign up and create a key. Copy the whole key (starts with sk-corr-) — you will paste it in step 3.

  2. 2

    Launch DeepSeek Harness

    Requires Node.js. One command — the web UI opens in your browser (127.0.0.1:3080):

    bash
    npx @deepseek-ai/dsh web
  3. 3

    Add Sozdai as a custom provider

    Open Settings → Models → Add a custom provider, and fill the form exactly like this:

    Provider IDsozdai
    Display nameSozdai AI
    Base URLhttps://api.foxwire.ai/v1
    API protocolopenai-completions
    API keyyour key from step 1 (sk-corr-…)
    Custom provider form in DeepSeek Harness

    Screenshot: official DeepSeek Harness docs (MIT)

    Then press “Fetch available models” — the model list loads automatically. Tick the ones you want (e.g. deepseek-v4-flash) and Save. Done.

Best models for DeepSeek Harness

Prices per 1M input tokens. All appear after “Fetch available models”.

Model idBest forContextInput / 1M
deepseek-v4-proFlagship reasoning1M104 ₽
deepseek-v4-flashFast & cheap default1M34 ₽
claude-sonnet-5Coding & agents1M480 ₽
kimi-k2.6Agent tasks, tools262K115 ₽
qwen3-coder-nextCheap coding specialist262K24 ₽
Optional: enable image input (official settings.yaml syntax)
yaml
llm-pi-ai:
  providers:
    sozdai:
      api: openai-completions
      baseURL: https://api.foxwire.ai/v1
      models:
        - id: gemini-3.7-flash
          input: [text, image]

Full price list.

FAQ

Why does the official DeepSeek API give 403 from Russia?

DeepSeek blocks Russian IPs and doesn't accept Russian cards. Sozdai provides direct programmatic access to the same V4 models with ruble billing.

Do I need a DeepSeek account?

No — only a Sozdai account and API key.

Can I pay with a Russian card?

Yes — top up your balance and pay per token, no subscription.

“Fetch available models” returned nothing — what now?

Check the Base URL ends with /v1 and the key starts with sk-corr-. You can also type model ids by hand — use the table above.

Which models besides DeepSeek?

26 models work with the same key: Claude, GPT, Gemini, Grok, Kimi, Qwen, GLM. Select them in the same provider after Fetch.

How do I enable image input for a model?

Per the official docs, hand-entered models are text-only by default. Add input: [text, image] for that model in $DSH_HOME/settings.yaml — example below the model table.

Also coding in the terminal? Same key powers Claude Code / OpenCode

Start in 3 minutes

Create a key, launch the harness, add the provider. Pay only for what you use.