cuelo start runs a ready-made speech-to-speech endpoint with the recommended stack. Clients prompt it; no project needed.

cuelo start runs the built-in endpoint (its agent name is cuelo) with the dashboard on the same port. It has no fixed persona or tools. Clients set instructions, tools and voice with OpenAI Realtime session.update, so one server can be any agent. See Prompts.

cuelo start                 # http://127.0.0.1:4141, opens the Talk page
cuelo start --no-open       # don't open a browser
cuelo start --port 5000

It reads .env from the current directory. cuelo-serve with no agent file runs the same agent without the browser step.

The stack

Part	Default
STT	Smallest, English
LLM	The first provider you have a key for: Cerebras `qwen-3.8-27b`, then Anthropic `claude-sonnet-5-5`, OpenAI `gpt-5-mini`, OpenRouter
TTS	Smallest, voice `aarini`
End of turn	turn-1-mini
Interruptions	MaAI (falls back to the backchannel detector without `cuelo[maai]`)
Backchannels	MaAI, with a 1.2 s hold after a backchannel (falls back to the VAD rule)
Store	`sqlite://cuelo.db` in the current directory

Changing it

Every part has an environment variable. Put them in .env:

Variable	Default	Values
`CUELO_STT`	`smallest`	any STT provider name (see Providers)
`CUELO_TTS`	`smallest`	any TTS provider name
`CUELO_VOICE`	`aarini`	provider voice ID. Clients may override it per session.
`CUELO_LLM`	`auto`	`auto`, `cerebras`, `anthropic`, `openai`, `openrouter`
`CUELO_MODEL`	per provider	model ID
`CUELO_TURNS`	`turn-1-mini`	`vad`, `smart-turn`, `turn-1-mini`, `maai`
`CUELO_BARGE_IN`	`maai`	`vad`, `detector`, `maai`
`CUELO_AGENT_BC`	`maai`	`vad`, `maai`
`CUELO_BC_HOLD`	`1.2`	seconds to wait after a backchannel before replying. `0` turns it off.
`CUELO_DB`	`sqlite://cuelo.db`	store URL
`CUELO_API_KEY`	unset	required on every request when set. Comma-separate several.
`CUELO_CORS`	unset	browser origins allowed to call the API, comma-separated
`CUELO_WORKERS`	`1`	server processes (`cuelo-serve`)
`CUELO_TORCH_THREADS`	`1`	PyTorch threads per process for the MaAI models

For example, a light stack that fits about 60 calls in one process:

.env

CUELO_BARGE_IN=detector
CUELO_AGENT_BC=vad

Prompting it from a client

Any Realtime client works. Send the persona with session.update. To make the agent speak first, send response.create:

{
  "type": "session.update",
  "session": {
    "instructions": "You are Maya, a friendly receptionist at a dental clinic.",
    "audio": { "output": { "voice": "aarini" } },
    "tools": [
      {
        "type": "function",
        "name": "book_slot",
        "description": "Book an appointment slot.",
        "parameters": { "type": "object", "properties": { "time": { "type": "string" } } }
      }
    ]
  }
}

{ "type": "response.create" }

Tools declared by the client run in the client. The agent sends a function_call item and waits for your function_call_output. See Realtime API.

The dashboard's Talk page does exactly this, with an editor for the prompt, tools and voice, and presets to start from.

Default endpoint

The stack

Changing it

Prompting it from a client

On this page

Default endpoint

The stackCopy Anchor Link

Changing itCopy Anchor Link

Prompting it from a clientCopy Anchor Link

On this page

The stack

Changing it

Prompting it from a client