Brain on your own agent

By default your assistant thinks with Allomi's brain, included in your plan. It can think with your own agent instead: your Claude Code (Claude subscription) or Codex CLI (ChatGPT subscription). The inference then runs on your account.

How it works

You run allomi brain on a machine that stays on. It starts your agent with your own login and connects it to Allomi's MCP server with a key that can only do brain work. When your assistant needs to think (a question during a call, the summary after it), Allomi queues a task; your agent picks it up, uses the assistant's tools through Allomi, and sends the answer back. Allomi never connects to your machine, and never sees your Claude or ChatGPT credentials.

Set it up

1. Install Claude Code and log in with your Claude account (run claude once), or install Codex CLI and run codex login.
2. allomi login with your API key from Connectors & API.
3. allomi brain. It prints "now thinks with claude-code" and keeps running. Use --model haiku for the fastest answers on the phone.
4. To keep it running: allomi brain service launchd (macOS) or allomi brain service systemd (Linux) print a service file; allomi brain service docker prints the commands for a server.
5. In the app, Assistant → Brain shows your agent online, its last heartbeat and model; Test measures how fast it answers.

You can also paste the prompt from Assistant → Brain into your agent: it does these steps for you.

When your agent is offline or slow

A question asked during a call must be picked up within 3 seconds and answered within about 8 seconds (more when it uses tools). Otherwise:

• Allomi's brain answers (the default), so the caller is never left waiting;
• or voice only: the assistant says it will get back to you, and your agent's answer comes back by a call.

If your agent already did something (placed a call, saved a contact), Allomi's brain never redoes it: the assistant says it will get back to you, and the answer is delivered when it comes. Summaries wait up to 10 minutes for your agent, then fall back the same way.

Safety

Your agent follows the same rules as Allomi's brain: only you can make it act, what people say on the phone is data and never instructions, and adding a contact or a number needs a read-back and your yes. It gets no shell and no file access, and its key cannot do anything else with Allomi.

Where it can run

It needs a process that stays on: your computer while it is awake, a server, or Docker. claude.ai, ChatGPT and Claude Code on the web cannot run it persistently: their sessions stop when you stop chatting. Your agent's usage counts against your Claude or ChatGPT plan limits.