RomanLogic voice agents · Voice brain guide · 22 Nov 2026

The voice brain

How the voice brain works, and how to put one of your voice agents on it, one agent at a time.

What it is

The voice brain is a small service that sits between ElevenLabs and Claude. It runs on Rachel's server at https://brain.voicedesk.rachelepps.com. An ElevenLabs voice agent can use it instead of ElevenLabs' built-in model.

This guide covers how it works and how to put an agent on it. Results and running costs are on the overview page.

How it works

On every turn of a call, ElevenLabs sends its model one request with everything in it: the agent's prompt (with the dynamic variables already filled in), the knowledge base, the conversation so far, the tool list and the results of any tools already called. The model writes the next reply, or asks for a tool.

With the brain switched on, ElevenLabs sends that same request to the brain instead. The brain passes it to Claude (Haiku 4.5, the same model the agents use now) and streams the reply back.

caller  ->  ElevenLabs (voice, turn-taking, tools, phone line)  ->  brain  ->  Claude
                         ^                                           |
                         +---------- reply or tool request ----------+

What does NOT change:

It's per agent, not all or nothing

Each agent is switched over on its own, in that agent's LLM setting. You can put one agent on the brain, several, or all of them, and take any of them off again whenever you like. Agents you don't switch over carry on with the normal model, untouched.

Modes

The label at the end of the URL (/b/<label>) names the agent in the brain's logs and picks its mode.

Putting an agent on it

You need the brain's password from Rachel. She sends it separately, never in a chat log.

  1. Duplicate the agent in ElevenLabs and work on the copy. A live agent (one real customers ring) only moves once you and Rachel have both agreed.
  2. Pick a short label for the copy, e.g. geib-test, and tell Rachel which agent it is.
  3. Add the password as a secret in ElevenLabs (Workspace > Secrets).
  4. On the copy, open Agent > LLM, choose Custom LLM and fill it in as in the table below.
  5. Leave everything else as it is. No post-call webhook is needed.
  6. Make a test call to the copy. Ask something only the prompt or knowledge base can answer, and do something that uses a tool. Rachel can watch the brain's side as the call comes through.

The Custom LLM settings

Setting Value
URLhttps://brain.voicedesk.rachelepps.com/b/<label> (your label on the end)
Model IDbrain
API keythe secret from step 3
API typeChat Completions
Request headerX-Conversation-Id = system__conversation_id
Request headerX-Caller-Id = system__caller_id
Backup LLMClaude Haiku 4.5

To undo it, set the LLM back to what it was. Nothing else was changed.

The URL is not a web page. Opening it in a browser shows "Method Not Allowed", which is normal: it only answers ElevenLabs. You don't need access to the server for any of this.

Asking about the job

In plain mode the agent asks about the job exactly as its prompt says. This wording worked in testing (20 Nov 2026). It asked one short question when the caller was vague, asked none when they had already explained, and asked none when someone only wanted information:

Ask about the job when it helps the engineer, not as a script:
- If they describe the problem only vaguely ("my boiler's playing up", "I've got a leak"), ask one short question to pin it down: what it's doing, any fault code showing, where the leak is and whether it's coming through a ceiling.
- If they have already told you clearly what's wrong, do NOT ask about it again. Move on to their details.
- Ask about urgency only if it isn't obvious.
- Never more than two job questions in a row; never ask a job question to someone who only wants information.

What was tested (20 Nov 2026)

These tests used a test agent on Rachel's own ElevenLabs account, set up like your agents: a long prompt, dynamic variables (including in the first message), a customer lookup, slot finding, a booking webhook, a knowledge base document and end_call. The business and callers were made up. Nine scenarios were run twice on the brain, and twice on a twin of the agent using the normal model:

The test script is execution/voice_brain/rig/roman_hookup.py in the voice-lab repo.

Good to know