Agent Routing Now Supports Custom Providers

Starting with MadoHub beta 0.3, you can connect any OpenAI-compatible API endpoint as a custom provider. This means you can run self-hosted models, fine-tuned variants, and specialized inference services alongside commercial APIs — all on the same canvas.

Why custom providers matter

Different tasks benefit from different models. A code generation tile might use a large, capable model, while a classification tile in the routing layer only needs a fast, small model. Forcing all tiles to use the same provider wastes either quality or cost.

Custom provider support lets you optimize each tile independently:

• Use a commercial API for tasks that require strong reasoning
• Use a self-hosted model for tasks involving sensitive data that can't leave your network
• Use a fine-tuned model for domain-specific tasks where a general model underperforms
• Use a fast, cheap model for transform nodes that only need to extract or reformat data

How it works

Open the provider settings in MadoHub and add a new custom provider. You need three pieces of information:

1. Endpoint URL — the base URL of the OpenAI-compatible API
2. API key — your authentication token (stored locally in the system keychain)
3. Model identifier — the model name to pass in API requests

Once configured, the custom provider appears in the model selector for any tile. You can mix providers freely — one tile running Claude, another running a local Llama model, a third using a fine-tuned GPT variant.

Routing between providers

The routing layer doesn't care which provider powers each tile. Connections work identically regardless of the underlying model. This means you can build workflows that leverage the strengths of different models at different stages.

A practical example: use a strong reasoning model to analyze a complex requirement, route the structured output to a fast code generation model, then route the generated code to a specialized review model fine-tuned on your codebase.

Supported API formats

MadoHub supports the OpenAI chat completions API format. Any service that implements the /v1/chat/completions endpoint with standard request and response schemas will work as a custom provider. This includes:

• OpenAI and Azure OpenAI
• Anthropic (via compatible proxy)
• Local inference servers (llama.cpp, vLLM, Ollama)
• Custom fine-tuned model endpoints
• Any OpenRouter-compatible service

Available now

Custom provider support is available in MadoHub beta 0.3, which you can download from the download page. Update your existing installation or grab a fresh copy.

If you run into compatibility issues with a specific provider, please file an issue on our GitHub repository with the provider details and we'll work on adding support.