Direct REST
Native HTTP, JSON, server-sent events, and WebSocket routes.
Tokaroo integrations
Bring your application or harness. Tokaroo handles the model route.
Use one of four compatible API formats for almost any server application, or start from a tested profile for eleven agent harnesses. Tokaroo does not take over your tools, files, or runtime.
Application formats
Most integrations are a base URL, key, and model change.
Use the native wire your application already understands. LangChain and the Vercel AI SDK can call these formats as optional client libraries; neither sits in Tokaroo's routing path.
OpenAI-compatible
Use existing OpenAI clients with a base URL and key change.
Anthropic-compatible
Messages, tool use, streaming, and token counting for Claude-oriented clients.
Gemini-compatible
GenerateContent, streamGenerateContent, and countTokens.
A harness does not need a named profile to call Tokaroo. Any trusted runtime that can send one of these compatible formats can use
max, pro, auto, fast, or basic.Eleven harness profiles
Use a tested path for the agent runtime you already operate.
Profiles preserve a harness's strongest native wire and add a bounded model-family affinity when compatibility testing shows that its tool loop performs better there.
Verified
Codex
Keep Codex's native Responses tool loop while Tokaroo selects and meters the model route.
Verified
Claude Code
Point Claude Code at Tokaroo's Messages surface without translating its native tool protocol.
Verified
Hermes Agent
Run Hermes through one Tokaroo endpoint with optimized model selection and fallback.
Conditional
OpenClaw
Use Tokaroo as an OpenAI-compatible provider for OpenClaw agents and skills.
Verified
Kimi Code
Connect Moonshot's coding harness to Tokaroo while retaining access to every compatible route.
Verified
DeepSeek Harness
Run the DeepSeek harness against Tokaroo rather than tying it to one upstream model account.
Verified
Qwen Code
Route Qwen Code through Tokaroo with a Qwen-aware preference and cross-family fallback.
Verified
Gemini CLI
Use Gemini CLI's native GenerateContent wire through Tokaroo instead of an OpenAI translation layer.
Verified
Grok Build
Connect Grok Build to Tokaroo with xAI affinity and compatible fallback routes.
Verified
Goose
Use Goose with Tokaroo as a custom OpenAI-compatible provider for model access and fallback.
In evaluation
OpenCode
Evaluate OpenCode through an inline Tokaroo provider before using it for production workloads.
Harness-aware, model-open
Native affinity improves the first choice without creating lock-in.
Codex often works best on OpenAI-family Responses routes; Claude Code on Anthropic Messages; Gemini CLI on Google GenerateContent; and Kimi, DeepSeek, Qwen, and Grok harnesses often benefit from their native model families.
Compatibility first
The request's wire, tools, output shape, context, and modality determine which models are eligible.
Quality band second
The requested Tokaroo mode establishes the task-specific quality band before family preference is considered.
Affinity, then optimization
A proven native family can lead inside that band, while a materially healthier compatible fallback remains available.
Model and provider families available to the router
GPT / OpenAIClaude / AnthropicGemini / GoogleGroqKimi / MoonshotDeepSeekQwen / AlibabaGrok / xAIOpenRouterMistralLlamaCohereAugure AIInterfaze AIRaspy multimodal
Harness and model remain separate choices. DeepSeek Harness, for example, can use any Tokaroo-compatible model; DeepSeek affinity is not a hard lock.
Reference implementations
Gwen proves the full harness path. Raspy proves multimodal routing.
They are integrations around Tokaroo, not extra harness profiles in the directory.
Gwen
Gwen runs ten active agent engines and sends their model traffic through Tokaroo's native compatible wires. This validates the separation between harness execution and model routing.
Raspy
Raspy powers Tokaroo speech, transcription, realtime audio, and avatar routes. Image and video requests use the broader compatible model network.
Connect the runtime you already trust.
Start with its native API format and Auto. Add the harness profile header where the client supports custom headers.