Cicero

Appearance

Using Cicero ​

You have run setup, started Cicero and opened the page. This is what day-to-day use looks like.

What to say ​

Talk to it the way you would talk to a colleague at the next desk. Everyday questions ("what's broken on CI?", "what did the last commit change?") go to the front desk. With an agent front desk, the agent answers everything and can use its tools. With a model front desk, the model answers quickly and hands harder work to your agent (see Thinking harder below).

On the web page, click Start conversation first, then hold SPACE or the orb while you speak (push-to-talk), or switch to hands-free. You can also type into the text box; typing runs the same pipeline without the microphone. Details are in the web voice guide.

Interrupting ​

Talk over Cicero at any time and it stops speaking and listens ("barge-in"). A small local voice-activity model checks that it is actually speech first, so keyboard noise and music do not cut it off. A quick "mm-hmm", "yeah" or "right" just tells it you're following, so it keeps talking. Typing while it speaks interrupts it too. Within five minutes, "continue", "go on" or "as you were saying" resumes the interrupted reply from what you actually heard.

"Details" after a short reply ​

Long replies are shortened by default: Cicero speaks the first four sentences and one closing line, and the full text lands in the chat pane. With a helper model, that closing line is a one-line summary of the rest; without one, it is a generic "plus N more sentences". Say "details" (or "tell me more", "read the rest", "full version") to hear the rest. Set web_voice.tldr.enabled: false to hear every sentence.

Thinking harder ​

With a model front desk and an ACP agent chosen on the Agent step, these phrases send that turn to the agent instead of the front desk:

  • "think hard"
  • "think deeply"
  • "think carefully"
  • "think it through"

The agent and the front desk keep separate conversations, so this suits one-off deep questions rather than follow-ups in the middle of a thread. You can change the phrases with brain.escalate.triggers.

Privacy ​

The Privacy step wrote one of two modes into privacy.mode:

  • local: nothing leaves this machine unless you allowed it. The front desk, the helper and speech run here. The wizard offers a cloud coding agent or a hosted task board only after you allow that item (privacy.allow); Telegram, which you add by hand for now, needs telegram in that list or cicero doctor warns. What leaves with each: prompts and code go to the agent's company, message text to Telegram, task text to the board.
  • cloud: conversation may use cloud models. The front desk and agents may be cloud services. The helper and speech still run here, and Telegram and task boards still need their own allowance.

This is a declared policy, not a firewall. The wizard only offers what the policy allows, and cicero doctor warns when the config drifts from it (for example, a cloud endpoint added by hand in local mode). Neither can see what a CLI agent does on the network once it runs.

Checking what is running ​

bash
cicero status          # the effective config and what is running right now
cicero doctor          # checks every configured part and prints fixes
cicero doctor --json   # the same checks as JSON, for scripts and AI agents

cicero doctor confirms binaries, ports and config, including the privacy warnings above. It cannot prove that an agent's CLI login works; one real spoken turn does.