Pick the model per agent, on your key
Dograh is the self-hostable voice agent platform with an open-source workflow builder. Each agent can run its own language model on a key in your name, and changing models is a settings change, so the call flow stays as you built it.
Save keys, assign models, compare calls
Store each provider once
Model configurationa named set of model settings
Under Models, save a configuration for each provider and key, then mark one as the Organization Default.
→ Name each one by the job it does
Give each agent its own model
Organization Defaultthe model agents inherit
A short booking agent can stay on a small model while a support agent with many tools runs a bigger one.
→ An API trigger can change the model for one call
Keep calls going when a model fails
LLM fallbacka backup model that steps in
Turn it on and a second model answers if the first errors or stays silent past 1.5 seconds, the default.
→ It stays off until you switch it on
Hosted, routed or on your own server
Claude runs through AWS Bedrock or OpenRouter, and Grok through OpenRouter
Each provider charges your account at its own published rate
A local model keeps the language step off every provider's servers
Test these before a live agent changes model
- Tool calls passThe smaller model books, looks up and hands over correctly on spoken input.
- Ten calls pricedToken counts from ten real call traces give the true cost of each call.
- Reasoning checkedThinking tokens bill as output, so turn reasoning down where the model allows.
- Fallback switched onA second model is set for the moments the first one errors.
- Cache not assumedBudget at the full rate until your own bills show a cache discount.
What a key in your name changes
Pay the provider directly
The model bill comes straight from the provider at its own rate, with nothing added on top.
Models change, flows stay
Swap the model in an agent's settings and its nodes, tools and prompts stay as built.
Go local whenever you choose
Point the same agent at Ollama or vLLM on your own server when a client asks for it.
Every turn in the trace
View Trace opens each call in Langfuse with the model and token counts for every AI step.
What you control when the key is yours
| What you get | Dograh | platforms that bundle the model |
|---|---|---|
| Your own key for the language model | ✓ | ✕ |
| A different model for each agent | ✓ | ✕ |
| Open-weight models on your own server | ✓ | ✕ |
| No per-minute platform fee when self-hosted | ✓ | ✕ |
Price every model family, call by call
Which LLM providers can I use with my own key?+×
OpenAI, Google AI Studio and Vertex, Azure OpenAI, Groq, Atlas Cloud, Hugging Face and OpenRouter, plus local models through Ollama or vLLM.
Can I run Claude or Grok?+×
Yes. Claude runs through AWS Bedrock or OpenRouter and Grok's text models through OpenRouter, each billed at that route's own price.
Does the platform add anything to my model bill?+×
No. With your own key, the provider charges you its own rate, and nothing is added to that bill.
Is there a per-minute platform fee?+×
Self-hosted Dograh has no per-minute platform fee, but servers, model keys and phone minutes still cost money. Dograh Cloud charges a 1¢-a-minute platform fee.
Can a single call use a different model?+×
Yes. An API trigger can override the model for one call, which suits trying a new model on a handful of real calls.
What happens if my model provider fails mid-call?+×
With LLM fallback on, a backup model takes over when the main one errors or sends nothing within 1.5 seconds by default.
Does my own key keep call data private?+×
It changes who bills you. The provider still processes what it receives, so run an open-weight model locally to keep that step in-house.
Bring your key. Pick the model per agent.
Self-host in one command, or start on Dograh Cloud.
$ curl … start_docker.sh && ./start_docker.sh