Skip to content

Chat#

Chat talks to the deployments of your workspace and to those shared with it, as the console's Playground does, made for a phone: the answer appears as it is written, code comes with Copy, and the conversation stays on the phone.

You need permission to call deployments (deployments:call) or deployments shared with the workspace (shared-deployments:call); without either, the tab is not shown.

Start a chat#

  1. Open Chat and tap + (or New chat on Home).

  2. Tap the model's name at the top to pick one. Each says its state:

    State Means
    ready It answers now.
    asleep — wakes in ~30 s Scaled to zero: your first message starts it, and the answer comes once a replica is up. The time is an estimate.
    starting… A replica is starting.
    unavailable It failed or cannot serve; pick another.

    Models shared with the workspace are listed under Shared with this workspace, with whose they are. Each also says whether it sees photos and whether it thinks (a reasoning model).

  3. Write and tap ↑. On a keyboard, Enter sends and Shift+Enter starts a new line.

While it writes, ■ stops it; what it wrote so far is kept. A reasoning model's thoughts are folded above the answer (Thought). Under each answer: Copy, Read aloud, and on the last one Answer again. Code blocks have their own Copy. Answers are markdown; images a model links to on the web are shown as links, never loaded.

Photos, PDFs and text files#

Tap + in the message box:

Choose Goes to the model as Limits
Take a photo, Photos Images, made smaller on the phone (at most 1568 px on the longer side) — only to a model that sees photos Up to 8 MiB of photos in one message; a file up to 40 MiB
A PDF or a text file Its text, read on the phone and put before your question, with the file's name Up to 20 MiB; the first 60,000 characters of each file

A scanned PDF has no text to read: attach its pages as photos instead. Your message shows each file by name (Attached: report.pdf), not its text.

Voice#

  • Dictate: with the message box empty, tap the microphone. If the workspace has a speech-to-text deployment (a Whisper model), your recording goes to it — on your machines, like a message — and the text lands in the box; tap Done when you finish. Otherwise the app uses the browser's own speech recognition, which in Chrome and Safari is the browser maker's service.
  • Read aloud: under an answer, tap the speaker. The phone's own voice reads it, on the phone; code blocks are named, not spelled out.

The model's own tools#

A deployment can have tools of its own — Functions and the code interpreter, set up in Give a model tools. When the model calls one, the call appears in the answer as it runs, with its arguments and result (tap it). A call held for approval says so and waits: its approvers find it in their inbox.

Instructions and profiles#

⋯ → Instructions sets what the model is told before your messages in this chat ("Answer in Portuguese, briefly"). Save instructions you reuse as a profile (Save as a profile, or Settings → Chat profiles) and pick it when you start a chat. Profiles are kept on the phone.

Where chats are kept#

On the phone only, in the app's own storage, separately for each account that signs in there. Astralyx keeps no copy: a message goes from the phone through the console to the machine serving the model, as you, and the answer comes back the same way. Tokens count in the workspace's usage, as in the Playground.

  • ⋯ → Delete this chat, or ⋯ beside a chat in the list, removes it.
  • Settings → Delete chat history on this device removes every chat, profile and unsent message of yours on that phone.
  • Another phone has its own chats: they do not follow you between devices.

Offline#

Without a network you can still read your chats and write: a message waits under Waiting for a connection and is sent, in order, when the phone is back online (tap Send now to try at once). A refusal — no permission, a model that is gone — is shown in the chat and not sent again.

Troubleshooting#

You see What to do
No model to talk to in this workspace yet Deploy one (Quick start), ask for one to be shared with you, or switch workspace on Home.
This model does not see photos Pick a model marked sees photos.
… is unavailable The deployment failed: see Troubleshoot a deployment, or pick another.
The answer takes long to start The model was asleep: the first message waits for a replica to start.
has no text to read (a scan?) Attach the pages as photos to a model that sees photos.