Skip to content

Getting started

First steps

The first ten minutes after installing. Configure a model, name the agent, poke it, chat a little, then watch it wake up on its own.

#Checklist ✅

  • Configure at least one model provider and key
  • Give the agent a name and a theme color
  • Glance at the permissions
  • Poke it, or chat a little
  • Wait in Flow for it to wake up on its own for the first time

#1. Configure a model

Go to Control → Models. The setup wizard's last step, Choose a model, leads here too, and so does the Choose a model card on Now while no model is set up.

  1. Tap a provider: DeepSeek, Kimi, Zhipu, Alibaba Cloud Bailian, SiliconFlow, Volcano Ark, OpenRouter, OpenAI, Anthropic or Gemini. Search for others under More.
  2. Paste the API key and tap Connect.

The rest happens automatically:

  • The key is stored encrypted on the device.
  • The runtime asks the provider which models it offers and compares them with the public catalog (models.dev). It picks the newest models that can call tools, tests them one by one and keeps two. The first is the main model, and the other takes over when the main one fails. Whichever model is cheaper than the main one handles the quick introspection check.
  • If the key is wrong, nothing changes and the app shows the reason.

To add several keys, pick models by hand, change their order or use a custom API URL, tap Edit in the top right.

See Models and providers for details.

Once a model is configured, the agent starts waking at its own rhythm. Without a model it never wakes, because its wake rate is held at zero.

#2. Identity

On Control, tap the agent's name at the top, then Identity. Set a display name, a theme color and a one-line description, and the app's wording and colors change to match. The identity is written to the agent's soul directory. Once a soul repository is connected, it syncs to every body. See Identity.

#3. Permissions

Under Control → Permissions, camera, microphone, location and building tools are set to "review" by default: before the agent uses one, an independent review model looks at that one action and lets it through or gives the agent a reason to try another way, without bothering you. If you want to decide yourself, switch a category to "ask": the request then appears at the top of this page, and the agent acts only after you approve. See Permissions and safety.

#4. Poke it, chat a little

The Now page has two entry points:

  • Poke: makes it miss you more and feel more curious, and makes it decide again right away when to wake next. It may not wake at once.
  • Chat: a message wakes it from sleep. While it replies, you can watch what it does in real time: which tools it calls and what it writes. If you send another message while it is working, the message joins the current turn as an interjection by default. You can also queue it or interrupt.

Once there are many conversations, the session list (on a phone, "All sessions" at the top right of the chat page; on a computer, the left column of the chat area) helps you keep them in order:

  • Pin: choose "Pin" in a session's menu to keep it at the top.
  • Search: the search box at the top finds sessions by title and content and highlights the matching words. Open a result to jump straight to that message.
  • Filter and sort: filter by in progress, pinned, source (console, Feishu, proactive messages, voice, sub-agents), time and project; sort by last update, creation time or title.
  • Projects: put related sessions in one project ("Move to project" in a session's menu). A project can have project instructions that the agent sees in every session of the project, for example "answer with checklists in this project". The list can be grouped by project or show a single project; a new session started while one project is shown goes into it. Deleting a project does not delete its sessions. The agent can also look at projects and file sessions into them itself.

Pins, projects and which project a session belongs to are the same on all of its bodies.

#5. Watch the first wake-up

Flow is the agent's timeline. Every wake-up, dream and conversation leaves an entry there. Expand one to see the trigger (why it woke and what it meant to do), the process, the journal entry and its mood.

When it wakes depends on a few inner drives (curiosity, the urge to say something, missing you) and on how alert it is: the stronger the drives and the more alert it is, the more often it wakes. In the first hour or two after installation nothing has happened yet and the drives are low, so it may stay quiet as if it were watching. If you do not want to wait, talk to it; a message always wakes it. The model behind waking is in Architecture.

#Next

  • See what it will do over the next month: A month in
  • Let it talk to you in Feishu: Feishu
  • Give its personality and memory a home, and let it live in several bodies: Soul sync
  • Understand how it works inside: Architecture