onekey-models

OneKey-Models

One-click access to all 108 latest models on TokenDance
inside DeepSeek Harness

English · 简体中文

Install · Features · What is TokenDance? · How it works · Testing · Video

---

API support provided by TokenDance. This plugin is an independent client of the TokenDance API platform; it is not affiliated with or endorsed by TokenDance.

Install

Clone the repository, then install it as a DSH bundle:

git clone https://github.com/du-yuxuan/OneKey-Models.git

In the DSH chat, ask the agent to install it:

install_bundle target=/absolute/path/to/OneKey-Models

Or with npm:

npm install onekey-models
node <dsh>/lib/bin.js plugin --profile <profile-name> add ./node_modules/onekey-models

Restart DeepSeek Harness (or the profile session) after installing. The settings page appears under Settings → Plugins → OneKey-Models.

Requires Node.js ≥ 22.19.0. Distributed through the DSH plugin markets, install_bundle, npm (onekey-models) and this GitHub repository — the same bundle from every channel.

Features

	Feature	What it does
1	API key management	OAuth (PKCE) sign-in opens the TokenDance authorization page; the returned key is stored locally in DSH's credentials store and never appears in source code or logs.
2	Model list configuration	Fetches the live catalog from https://tokendance.space/gateway/v1/models, lets you select multiple models, and choose which ones show up in the model picker (empty selection = show everything).
3	Special model configuration	Image-generation models are wired to their matching application (openai:image-generations / ark:image-generations); Jev judgment models are wired to the typesafe:systemone protocol — plus video, speech, audio, music, search, embedding and rerank models.
4	Core capabilities	Chat, image generation, video generation and multimodal (text + image) input work out of the box, exposed both to the model picker and as five agent tools (tokendance_models, tokendance_chat, tokendance_image, tokendance_video, tokendance_jev).

Enrichment data (logos, pricing, context windows, release dates, modalities) is synced automatically from models.dev.

Catalog at a glance

catalog: 108 models from https://tokendance.space/gateway/v1/models
  jev 2 · image 3 · video 18 · embedding 2 · rerank 1
  speech 3 · audio 3 · music 2 · search 3
  routable 71 models
✓ catalog check passed (108 models, 71 routable)

routable = models whose protocol maps onto a DSH wire protocol (openai-completions / openai-responses / anthropic-messages) and therefore appear in the model picker.

What is TokenDance?

TokenDance is an AI model aggregation gateway — think of it as one front door to many large models:

• 1 key, 1 base URL — sign up once, get one API key, point any compatible SDK at one address.
• 108 of the latest, hottest models — GLM, DeepSeek, Kimi, Qwen, MiniMax, Seedance, Doubao and more, kept up to date.
• Pay as you go — no per-vendor accounts, no juggling separate keys and billing pages.

New to all this? The short version: instead of registering with a dozen AI vendors and copying a dozen keys, you register once with TokenDance and this plugin wires every model straight into DeepSeek Harness's model picker.

This plugin does the wiring for you: install → sign in once → pick your models → they appear in the picker.

How it works

┌─────────────────┐   OAuth (PKCE)    ┌──────────────────────┐
│ DeepSeek Harness │ ───────────────▶ │ tokendance.space      │
│  OneKey-Models   │ ◀─────────────── │  authorization page   │
└────────┬────────┘   API key (local) └──────────────────────┘
         │                                          ▲
         │  write provider block                    │ GET /gateway/v1/models
         ▼                                          │ (live catalog)
┌─────────────────┐                                 │
│   llm-pi-ai      │ ────────────────────────────────┘
│   providers.     │
│   tokendance     │ ──▶ model picker: 108 models, 71 routable
└─────────────────┘

1. Sign in — the plugin opens the TokenDance authorization URL, receives a one-time code on a loopback callback, exchanges it for an API key, and stores it via DSH's credentials service (~/.dsh/.credentials.yaml, mode 600).
2. Pick models — the live catalog is fetched, enriched with models.dev metadata, and written into the llm-pi-ai settings namespace as a provider block.
3. Use them — selected models show up in the model picker; image and Jev models additionally get dedicated tools, and any picture or video can be requested from a chat prompt.

All local HTTP endpoints are guarded by a loopback + sec-fetch-site trust fence (inbound requests must come from the local origin, never a cross-site page).

Testing

npm test              # 71 unit tests (node:test)
npm run check:catalog # live catalog acceptance (108 models, 71 routable)

The unit suite covers config normalization, catalog mapping, the settings patch whitelist, Jev question validation, the five video protocols' submit-and-poll shapes, PKCE/OAuth URL construction, the trust fence, credential helpers, and the HTTP layer (exercised against a local stub gateway — no real key required).

Video

90-second intro (1920×1080, 30 fps) in nine scenes: the pain of a dozen keys, the gateway, the reveal, the three steps, the agent tools (image + video), Jev decisions, the install command, and the TokenDance credit. Rebuild it with npm run render inside video/.

Plugin page

The settings page shows the “API support provided by TokenDance” credit, the TokenDance brand mark, and a plain-language introduction of what TokenDance is. Sign-in, the balance card, redeem/top-up (as centered dialogs) and the model list all live on that one page.

Agent tools

Five tools are registered into the DSH agent registry, so a chat prompt such as “generate a video of a whale swimming” becomes a real call:

Tool	What it does
tokendance_models	Lists the live catalog, optionally filtered by capability (image, video, chat…)
tokendance_chat	One message to one model, with optional images attached
tokendance_image	Generates an image and returns it straight into the conversation
tokendance_video	Submits a video task and polls it; resumable by taskId
tokendance_jev	Asks a Jev model for typed decisions (yes/no + probability, choice, score)

A companion skill, tokendance-media (installed to ~/.agents/skills/), teaches the agent which tool to reach for and how to read the pending / taskId outcomes. Each tool has its own on/off switch in the plugin's advanced settings.

Repository layout

index.js            Host half: settings, OAuth, catalog, HTTP routes, tools
client.js           Client half: settings page (slot + locale registration)
src/                config, catalog, credentials, http, image, jev, oauth,
                    patch, pi-ai, tools, trust, video (submit-and-poll for
                    seedance / kling / minimax / wan3 / happyhorse)
locale/             zh + en metadata
test/               node:test suites
scripts/            check-catalog.mjs, oauth-once.mjs
logo/               brand assets
video/              Remotion project
out/                intro.mp4 + poster

License

MIT. “TokenDance” is a trademark of its respective owner; this project is not affiliated with or endorsed by TokenDance. The API access it configures is provided by TokenDance.