DeepSeek Harness

Point DeepSeek's open-source agent harness (dsh) at Tokens as a custom OpenAI-compatible provider, keep the key in an environment variable, and run DeepSeek models — or the rest of the catalog — through the agent loop.

DeepSeek Harness is DeepSeek AI's open-source agent runtime, invoked as dsh (released August 2026). The model does the thinking; the harness runs the loop around it — tools, shell, files, sessions, sandboxing and context — with almost everything wired up as replaceable Cordis plugins. It ships a CLI, a TUI and a Web UI, plus Standard, Code, Minimal and Creator modes.

dsh runs any OpenAI-compatible endpoint, so it points at Tokens with a config change and no extra plugins. Create a key at your dashboard if you don't have one. API keys explains spend caps and model allow-lists.

Install dsh#

npm install -g @deepseek-ai/dsh

npm install -g @deepseek-ai/dsh

That installs the dsh command. dsh web starts the Web UI on your machine. Your settings live in $DSH_HOME/settings.yaml (default ~/.dsh/settings.yaml) — the same file the Web UI writes.

Option A: the Web UI#

The fastest path. Run dsh web, open the UI, and go to Settings → Models:

1. Click Add custom provider.
2. Fill in the form:
   • Provider ID: tokens (lowercase, stable)
   • Display name: Tokens
   • Base URL: https://tokens.bd/v1
   • API protocol: openai-completions
   • API key: your tok_live_... key
3. Click Fetch available models to pull Tokens' model list, or add a model manually with its full id (e.g. deepseek/deepseek-v4.1-flash).
4. Save, then pick the provider and model in the chat tab's model picker.

Option B: edit $DSH_HOME/settings.yaml#

Point the built-in DeepSeek adapter at Tokens. No code changes needed:

llm-deepseek:
  baseURL: https://tokens.bd/v1
  apiKeyEnv: TOKENS_API_KEY
  models:
    - id: deepseek/deepseek-v4.1-flash
      name: DeepSeek V4.1 Flash (via Tokens)
      contextWindow: 1000000
agent-default-model:
  provider: deepseek-official
  model: deepseek/deepseek-v4.1-flash

Three points to get right:

• baseURL is the gateway root. /chat/completions is appended automatically, so end it at /v1 — not at /v1/chat/completions.
• The model id must carry the provider/ prefix. Tokens routes on the full id (deepseek/deepseek-v4.1-flash); the bare name is rejected. Copy ids from /models or GET /v1/models.
• The key resolves from the environment named by apiKeyEnv on every request. Export it in the shell that starts dsh:

export TOKENS_API_KEY="tok_live_your_key"

$env:TOKENS_API_KEY="tok_live_your_key"

Changes take effect on the next request — no restart needed. To go back to the official DeepSeek endpoint, edit settings.yaml again or pass --model / --provider for a single session.

Switch models#

Change models[].id and agent-default-model.model to another id from /models or GET /v1/models, and update contextWindow to the model's real window (DeepSeek V4/V4.1 Flash models carry 1M). Tokens isn't limited to DeepSeek models: any id in the catalog works through the same adapter, since the gateway translates the request for the upstream provider.

Inside the CLI, /model <model> switches the model for the current session.

Choosing a model helps you pick an agent-suitable model: look for tool calling support and a context window your workload fits in.

Verify it works#

dsh "Reply with the single word OK"

It should answer once and exit. If it starts looping or its tool calls fail, the model likely lacks OpenAI-style tool calling — switch to another model.

Troubleshooting#

401 or MISSING_CREDENTIAL. The env var named in apiKeyEnv isn't visible to dsh. Check the spelling and that it was exported in the shell that launched the harness.

Model not found / gateway rejects the request. The model id is missing its deepseek/ (or other provider) prefix. Use the full id from /models.

Context window looks wrong. Set contextWindow explicitly in settings.yaml — dsh trusts your config first. Copy the real value from /models.

403 model_not_allowed_on_key or 429 window_exhausted. These are key and plan limits, not harness errors. See Error code reference for every error code.

Web search inside dsh. The built-in web_search tool speaks the Anthropic-compatible Messages API. Point it at Tokens' Anthropic-compatible endpoint (https://tokens.bd — SDKs append /v1/messages), following the base-URL rules in Authentication.