operational/41 regions

One key for the whole model pool.

Skidder sits in front of every frontier lab with one endpoint, one sliding-window limiter and one bill. Change the model string, nothing else moves.

Get your key Quickstart

skidder · gatewaycopy

# unified gateway — one call, pooled frontier inference
curl https://api.skidder.wtf/v1/chat/completions \
  -H "Authorization: Bearer sk_live_9f2a•••c41b" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-sonnet-5","stream":true}'

claude-sonnet-5

41 regions

latency

input cache5.5ms

output24ms

ttft120ms

first token12.5t/s

wait202ms

one endpoint

Claude, GPT, Gemini, Kimi.

Every model speaks the same wire format. Point an OpenAI-compatible client at the base URL below and it just works.

https://api.skidder.wtf/v1

AnthropicOpenAIGooglexAIMoonshotAlibabaZhipuDeepSeekMetaMistralNVIDIA

everything in the pool

Claude Opus 5.1Claude Sonnet 5Claude Haiku 4.5GPT-6.1 SolGPT-6.1 TerraGPT-6.1 LunaGPT-5.5 CodexGemini 3.1 ProGemini 3.1 FlashGrok 4.20Grok 4.6Kimi K3Qwen 3.8 MaxGLM 5.2DeepSeek V4 ProLlama 4 405BMistral Large 3Nemotron 3 Super

why bother

Built for agents that
never stop running.

Long-lived agents do not tolerate provider churn. Upstream goes down, rate limits move, a model is deprecated mid-run. Skidder absorbs that so your process keeps its footing and keeps its keys.

Drop-in OpenAI shape

Streaming, tool calls, JSON mode and vision work the same on every model. Nothing to rewrite.

Real sliding windows

Rate limits count over a rolling interval, so bursts land where you expect instead of at the reset.

Per-key ceilings

Spend and RPM caps live on the key. A runaway loop kills itself, not your month.

One ledger

Every model, every key, one export. Reconcile in a spreadsheet or pipe it into your own stack.

failover.loghealthy