TokenDance

TokenDance is an LLM API aggregation gateway from the team behind the AI product community Watcha, self-described as "a Chinese OpenRouter", unifying access to 97 Chinese models from 16+ providers behind a single key. It packages multi-model access, intelligent routing, unified billing, and failover into one developer-friendly layer, mainly serving independent developers, agent teams, and early-stage AI startups in mainland China. Note: the same team also runs the Watcha AI product community — two products from the same people, not duplicate entries.

Core Capabilities

• Native three protocols: Natively supports OpenAI / Anthropic (Claude) / Google GenAI (Gemini) — official SDKs connect by changing only the base URL and key, unlike relays that only emulate OpenAI
• Smart routing and failover: Routes automatically by model name to the right provider, switches automatically among multiple endpoints of the same model, and supports candidate lists via the extended models:[...] field for stepped fallback
• Unified billing: Cross-provider token accounting and balance deduction in one place — one dashboard for all usage, no topping up vendor by vendor
• Rich model catalog: 97 models verified live via the API (the homepage's "8+" is stale), including MiniMax, Qwen, Kimi, Zhipu, DeepSeek, and Doubao
• Developer infrastructure: Playground, API key management, call logs, async webhooks, OAuth-style key authorization, and a balance-query Open API
• AI-facing docs: Provides llms.txt / llms-full.txt so Claude Code, Cursor, and similar tools can integrate on their own

Use Cases Independent developers, one-person companies, and early AI teams in mainland China, especially Claude Code / Codex / Cline users blocked by "OpenAI-compatible-only" relays, and agent developers needing low-cost multi-model switching with failover. Not for non-technical users who never write code (user reviews agree it is "of no use to ordinary users").

Unique Advantages Against the many OpenAI-compatible-only relays in China, native three-protocol support is a hard differentiator — Anthropic-format tools connect directly, rare among Chinese gateways. The background is also more verifiable than a typical "individual webmaster relay": the operator Watcha (Hangzhou Siyi Technology Co., Ltd.) has Hangzhou city government media coverage and a Sequoia China seed round, with underlying compute supplied by Tsinghua-affiliated Infinigence (over ¥2.2 billion raised to date) under a publicly declared strategic partnership; per-model prices are public and subsidies generous. Shortcomings are equally clear: almost no overseas models, and fewer models in absolute terms than OpenRouter or SiliconFlow.

Editor's Verdict The product is real — 97 working models, high-quality documentation engineering, and native three-protocol support beyond its peers, making it a legitimate choice as a domestic multi-model gateway. But the risks must be stated honestly: no terms of service or privacy policy anywhere on the site (the login page says "by logging in you agree to the terms of service" yet that link 404s, with no public terms on how data is handled or how keys are stored); the operating entity is not disclosed on the site (the Zhejiang ICP filing number 2024107375 points to Hangzhou Siyi Technology Co., Ltd.); the domain was registered only on 2026-03-25, roughly six months ago; official PR and changelog contradict each other on whether online top-ups are supported; some vendor logos are hot-linked directly from their CDNs; and independent third-party reviews are essentially zero — praise on Watcha comes from the same source as the vendor, a conflict of interest. Disambiguation: searching "TokenDance" also surfaces a UCLA arXiv paper and the GitHub org TokenDanceLab, both unrelated to this product. Use caution for sensitive data or strict compliance scenarios.