Max
The hardest and highest-stakes work. Tokaroo selects economically within the strongest measured task-specific quality band.
95% of ceiling fallback quality floor
One API. 714+ AI models.
Up to 95% less.*
Install one Tokaroo key in your backend or harness. Tokaroo connects the agent and application ecosystem to the live model network, selects the compatible provider and model, handles fallback, and meters the result.
Provider and model familiesGPT / OpenAI, Claude / Anthropic, Gemini / Google, Groq, Kimi / Moonshot, DeepSeek, Qwen / Alibaba, Grok / xAI, OpenRouter, Mistral, Llama, Cohere, Augure AI, Interfaze AI, Raspy multimodal, and more
HarnessesCodex, Claude Code, Hermes, OpenClaw, Kimi, DeepSeek, Qwen, Gemini CLI, Grok, Goose, OpenCode, and compatible apps
Optimized routingMax, Pro, Auto, Fast, Basic
Server-side and trusted harness use only. Never ship a permanent key in browser or mobile code.
Models tracked
714+
Measured savings
50.8%
Measured comparisons
38
Tokens routed
354.2M
One integration
Your software asks. Tokaroo decides where it runs.
Keep provider selection, fallback, billing, and routing operations out of your application code.
POST /v1/chat/completionsOpenAI-compatible
curl https://api.tokaroo.com/v1/chat/completions \
-H "Authorization: Bearer $TOKAROO_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "auto",
"messages": [
{"role": "user", "content": "Ship the fix"}
]
}' Request completed
Requestedauto
Tokens1,284
Tokaroo charge$0.0042
The response stays compatible with the format your application already expects. Tokaroo usage headers provide the requested mode, trace, tokens, latency, and charge.
Five virtual models
Choose the outcome, not the vendor catalog.
Quality ratios compare each candidate with the best freshly measured model for that task family. They are not statistical percentiles.
Pro
The best all-around mode for important technical, business, production, and customer-facing work. Most serious work belongs here.
85% of ceiling fallback quality floor
Auto
Automatically applies the capability each step needs by balancing quality, cost, speed, reliability, tools, and escalation risk.
50% of ceiling fallback quality floor
Fast
Rapid answers, edits, and back-and-forth. It keeps an acceptable task-specific quality bar, then prioritizes latency.
50% of ceiling fallback quality floor
Basic
Everyday questions and routine tasks where you need an answer, not a project. Coding and website builds are intentionally excluded.
No relative quality floor
These percentages are conservative fallback floors. When Tokaroo has enough fresh, diverse evaluations, task-specific quality clusters determine the qualifying band dynamically.
General chat and reasoning
Awaiting fresh evidence | Conservative fallback floors
- ceiling0 freshly tested0 decision-grade0 provider families
Max0measured candidates95% of ceiling fallback quality floor
Use what you already have
Four wire formats. Harness-ready configuration.
Tokaroo is the router, not the harness host. Your app or agent runtime keeps control of tools, files, and execution.
RESTNative HTTP and streaming
OpenAIChat and Responses
AnthropicMessages and tools
GeminiGenerateContent
Eleven harness profiles
Ten active Gwen engines plus OpenCode under compatibility evaluation. Each public profile is promoted only after its Tokaroo wire passes conformance checks.
Codex
Responses native | OpenAI-first
Claude Code
Messages native | Anthropic-first
Hermes
OpenAI-compatible | optimized routing
OpenClaw
OpenAI-compatible | conditional
Kimi
OpenAI-compatible | Moonshot-first
DeepSeek
OpenAI-compatible | DeepSeek-first
Qwen
OpenAI-compatible | Qwen-first
Gemini CLI
GenerateContent native | Google-first
Grok
OpenAI-compatible | xAI-first
Goose
OpenAI-compatible | optimized routing
OpenCode
OpenAI-compatible | in evaluation
Harness-aware routing preserves each tool's native wire and favors model families proven to handle its tool loop well. The preference applies only inside the selected quality band; verified cross-family fallback remains available unless you explicitly pin a model.
Gwen reference integration
Harness orchestration through one Tokaroo key.
Gwen chooses and runs the harness. Tokaroo supplies scoped runtime credentials, makes the optimized routing decision, enforces budgets, and records usage. That same pattern is available to other AI products.
User Gwen Harness Tokaroo Model
Raspy multimodal rail
Speech, realtime, and avatars through Tokaroo.
Tokaroo routes metered TTS, transcription, realtime voice, and avatar requests to Raspy, with compatible managed fallback where available. Customers keep the same account, key, budgets, and usage ledger.
App Tokaroo Raspy Voice or avatar
Keys belong in trusted runtimes.
AuthenticationUse environment variables or a secrets manager in your backend or harness. Browser and mobile clients should call your backend or use short-lived, scoped credentials.
One key. Five intelligent models. Every major provider.
Start with Auto. Choose Basic, Fast, Pro, or Max when the workload has a clear cost, latency, or capability requirement.