How to connect DeepSeek Harness to the Tokenator API

What you end up with

Once configured, DeepSeek Harness stops talking to its default provider and sends every request to Tokenator. Billing comes out of your key's token limit, and the list of available models is defined by the key rather than by the tool.

Requirements and supported operating systems

• Windows 10/11, macOS 12+, or a modern 64-bit Linux.
• Outbound HTTPS access to the Tokenator domain.
• A Tokenator API key that looks like sk-your-tokenator-key.

You need Node.js: the harness is started through npx. Building from source additionally needs pnpm. Everything is configured in the browser, and a DeepSeek account is not required for a third-party provider.

Installing DeepSeek Harness

npx @deepseek-ai/dsh web

The command starts the web UI at http://127.0.0.1:3080 and opens it in the browser. Pass --no-open to keep the server without opening a tab.

git clone https://github.com/deepseek-ai/deepseek-harness.git
cd deepseek-harness
pnpm install
pnpm run build
pnpm dsh web

Setting up Tokenator

1. Register in your account and buy a token bundle.
2. The key appears in the account right after the payment is confirmed — copy it whole, including the prefix.
3. The examples below use the placeholder sk-your-tokenator-key. Never publish a real key in repositories or screenshots.

Configuration

1. Open Settings → Models in the web UI.
2. Press Add a custom provider.
3. Provider ID in lowercase, for example tokenator. The id is permanent: a provider cannot be renamed, only replaced by a new one. The display name, address, protocol, credential and models stay editable.
4. Base URL: https://api.tokenator.top/v1
5. API protocol: openai-completions, which is OpenAI Chat Completions. A provider speaks exactly one protocol.
6. API key: sk-your-tokenator-key
7. Under Model catalog press Fetch available models: the request goes to GET https://api.tokenator.top/v1/models with the key currently in the form. Tick the models you want and press Add selected. IDs can also be typed by hand, for example gpt-5.5.
8. Save the provider and pick a model in the picker — the selected model becomes the default for new sessions.

llm-pi-ai:
  providers:
    tokenator:
      apiKeyEnv: TOKENATOR_API_KEY
      api: openai-completions
      baseURL: https://api.tokenator.top/v1
      models:
        - id: gpt-5.5

The same thing can be written straight into $DSH_HOME/settings.yaml — the very document the settings page writes. It also holds the fields the form has no place for: headers, timeouts, reasoning levels and compatibility switches. When the browser runs on the same machine as the server, Open configuration file in the settings header opens it.

Your first request

Before starting the tool it helps to confirm that the key and the model both work. A single curl request answers both questions.

curl https://api.tokenator.top/v1/chat/completions \
  -H "Authorization: Bearer sk-your-tokenator-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.5",
    "messages": [{"role": "user", "content": "ping"}]
  }'

Open http://127.0.0.1:3080, pick the model in the picker and send a message. A session that has already sent a request keeps the model recorded in its own log: the new model applies to new sessions.

A successful response means the key is valid, the model is allowed for it and the upstream is reachable. Usage appears in your account right after the request.

Switching models

Models are edited in the provider card: Fetch available models asks the endpoint for the list, and an ID typed by hand works exactly the same. Picking a model in the picker makes it the default for new sessions.

Examples of model IDs available right now:

• claude-haiku-4-5
• claude-opus-4-6
• claude-opus-4-7
• claude-sonnet-4-6
• gpt-5.5
• deepseek-v4-pro

Current IDs and specs live in the AI model catalog, which also shows which models are available right now. The model comparison helps choose between close options.

Configuring reasoning

A model typed in by hand declares no reasoning levels, so its Effort entry does not appear in the menu and the endpoint alone decides whether the model thinks. Levels are declared with reasoningEfforts in $DSH_HOME/settings.yaml: the key is the menu entry and the value is what travels on the wire as reasoning_effort. Only off may stay empty, because for most endpoints not thinking is the absence of the parameter.

llm-pi-ai:
  providers:
    tokenator:
      models:
        - id: gpt-5.5
          reasoningEfforts:
            off:
            high: high
            max: max

On top of that the key itself has a reasoning switch in your account, which works independently of the tool. Thinking is billed as output tokens.

Common errors and how to fix them

FAQ

Read next