Developer onboarding
Register, land directly in the console, and move from empty balance to first relay request with minimal friction.
Token relay platform
Operate a prepaid AI relay without making developers think about upstream routing.
Launch developer accounts, fund wallets, issue reveal-once API keys, and expose a clean public model catalog while keeping provider credentials inside your own platform.
Wallet-backed usage
Top up, meter relay traffic, and keep charges tied to public aliases.
Controlled access
Issue platform keys once, rotate them fast, and avoid exposing upstream secrets.
Operator visibility
Inspect users, route health, and billing adjustments from a dedicated console.
Billing clarity
Inspect wallet balance, top-up history, usage charges, and adjustments without hunting through raw logs.
Model access
Browse public aliases, pricing, and modality groupings while keeping upstream providers abstracted away.
One interface, 160+ models
Text, image, embeddings, video and speech all go through the same OpenAI-compatible interface, the same account and the same wallet. Switching models means changing one alias, not your code.
137text models
11image models
10embedding models
5video models
2speech models
Covers model families including DeepSeek, Kimi, MiniMax, Qwen, Doubao Seedream / Seedance, BGE embeddings and CosyVoice speech, with context windows up to 10M tokens.
Counts above are enabled models by modality. The catalogue moves as upstreams change; the copy that is always current is GET /v1/models.
Reliability: what we do, and what we do not yet promise
The real risk in a multi-model relay is not latency - it is one upstream misbehaving and taking the whole chain down with it. What we do about that:
- Ordered fallback chain: switching within an equivalent-capability group, so one upstream being down does not mean your request fails.
- Circuit breaking: upstreams that fail repeatedly are pulled out automatically instead of eating your retry budget.
- Balance gate: insufficient balance returns 402 outright - no overdraft, no silent deduction, the ledger reconciles.
We do not publish an availability SLA yet. Not because we cannot hold one, but because of a rule we set ourselves: availability gets measured before it gets promised, and upstream health snapshots are still accumulating. Publishing a number before that would ask you to plan production against a promise we have not verified.
Frequently asked
What is Model-as-a-Service (MaaS)?
Large models from several vendors placed behind one interface, one set of credentials and one wallet. You do not register, top up and adapt an SDK vendor by vendor, nor maintain a key and an invoice for each.
How is this different from calling each vendor's API directly?
Three problems go away: multiple keys and invoices, per-vendor interface differences, and writing your own fallback when one vendor breaks. One thing is gained: usage and cost for every call are visible in one place. The cost is one extra network hop.
How compatible is it?
An OpenAI-compatible interface - in most cases swapping base_url and api_key is enough, and tool calling (tools / tool_calls) is passed through intact. Capability differences for a specific model are as annotated in the catalogue.
How is it billed, and can we overspend?
Actual usage is deducted from the wallet, with usage and cost visible in the console. Insufficient balance returns 402 rather than running an overdraft - so there is no way to quietly accrue a debt.
Contact us
For business enquiries, custom configurations and quotes on compute and model access, reach us directly.