What Universe Bot can do
Can each agent use a different model and subscription?
Yes. Each agent has its own provider and model, and you can mix them within one team. To use a subscription, Universe Bot runs the official command-line tool that you installed and signed in to on your own computer. We never touch your account password.
| Connection | What you need | What it can do |
|---|---|---|
| Claude (Claude Code CLI) | A paid Claude subscription, or any sign-in that works with Claude Code | Chat, read and write the work folder, run commands, operate a browser, external MCP servers, image attachments, web search |
| ChatGPT (Codex CLI) | A paid ChatGPT subscription | Chat, read and write the work folder, run commands, image attachments, web search |
| Gemini (Gemini CLI) | A Google account | Chat, read the work folder, full access mode, web search (no image attachments) |
| OpenAI-compatible API (for example OpenRouter) | An API key of your own (free tiers work) | Chat, messaging teammates and team tools (cannot read or write files) |
Model shortcuts: Claude offers haiku, sonnet and opus; Gemini offers auto, pro, flash and flash-lite; Codex lists the models your account can use. If a model is not in the shortcuts, type its name.
What are delegation and the mailbox?
Delegation means agents send each other messages. Once the coordinator has passed work to a teammate it does not wait; it carries on with its own tasks. The teammate reports back when done, and the report wakes the coordinator up to continue.
- Every agent has its own work queue and handles incoming messages in order.
- The coordinator can hand work to several teammates at once, and they work in parallel.
- If a teammate forgets to report back, its last reply is delivered to whoever assigned the work, so the coordinator is never left waiting.
- Loop protection: when a message has been passed along too many times, it stops and tells you in the chat.
- You can make a teammate a commander that manages its own teammates, up to three levels deep.
- You can open the conversation between any two agents and step in with any agent at any time.
What are the usage panel and Clear session?
Every agent has its own usage panel that shows how many tokens it has used. Clear session starts that agent on a fresh conversation to save quota, while its memory, personality and rules stay the same.
| Item | Details |
|---|---|
| Time ranges | Current session, today, all time |
| Numbers | Turns, input tokens, output tokens, cache reads, cost (only for services that report cost) |
| Subscription quota | 5-hour and 7-day quota with reset times. This is per account and shared by every agent on the same provider (currently reported by Claude only) |
| Kept by Clear session | Memory, personality, rules, avatar and usage history; the old chat is archived and still viewable in history |
| Changed rules or memory | Take effect on the next turn, no Clear needed |
How do routines work? Do missed runs catch up?
A routine delivers a fixed message to a chosen agent at the time you set, straight into its usual conversation. If the app was not running at that time, the most recent missed run happens the next time it starts and is labelled as a catch-up with the original time.
| Item | Details |
|---|---|
| Frequency | Daily, weekly (pick several days) or monthly (day 1 to 31; months without that day run on their last day) |
| Time | 24-hour clock in 5-minute steps, in your computer's time zone |
| Catch-up | Only the most recent missed run, never a backlog |
| Limit | Up to 20 per agent |
| Who can create them | You on screen, or an agent you made a commander |
| Deleting | When an agent wants to delete a routine it shows a confirmation card in the chat; nothing is deleted until you press it |
| In the background | Closing the window keeps the app running in the system tray; to quit completely, choose Quit from the tray |
What is browser takeover?
When an agent operating a website reaches a step that needs a person, such as a sign-in or a verification check, it asks you to operate its browser right inside Universe Bot. When you are done, it picks up again.
| Item | Details |
|---|---|
| Available for | Claude CLI agents with the browser switch turned on in settings (off by default) |
| Browser | The Google Chrome already installed on your computer. Each agent gets its own separate profile with its own sign-ins, and your everyday Chrome data is never touched |
| What the agent can do | Open URLs, read the page structure, click, type, press keys, scroll, go back, use tabs, take screenshots |
| What it does not do | Download or upload files; open local files or other non-web addresses |
| Running at once | Up to 3 browsers, each closed automatically after 10 idle minutes |
| Taking over | The agent sends a card asking you to take over. Open it to see the live screen, click and type, then press Done or Skip |
⚠️ With the browser switch on, an agent can act with the accounts you are signed in to. Malicious content on a web page could mislead it, so only turn it on for tasks you trust.
Can I talk to the coordinator on Telegram or iMessage?
Yes. Once set up, you can message the coordinator and send it photos from Telegram or iMessage on your phone and get its replies there. It is the same conversation as on the desktop. iMessage runs through a third-party service, Photon (photon.codes).
| Item | Telegram | iMessage |
|---|---|---|
| Phones | Any phone that runs Telegram | iPhone |
| What you need | Your own bot, created with @BotFather | A Photon (photon.codes) account and project (Apple does not open iMessage to third parties, so it goes through Photon) |
| Pairing | The app shows a 6-digit code (valid for 10 minutes); send it to your bot | Same, sent to the number Photon gives you |
| Who you talk to | The coordinator only (it finds teammates for you) | Same |
| Photos | Yes | Yes. Linux cannot read the iPhone HEIC format, so switch your camera to Most Compatible or send screenshots; on a Mac they are converted automatically |
| Phone commands | /start for help, /clear to restart the coordinator's conversation | Same |
- Your computer and the app need to be running to receive messages.
- Only the paired phone can send commands.
- Your phone shows plain text only; anything that needs a button, such as confirming a deletion or signing in to a site, is left for you on the desktop.
- Messages pass through Telegram's or Photon's servers (never ours). See the privacy notice for details.
What are permission levels, the web switch and the sandbox?
Each agent has a permission level that decides what it can touch on your computer. Web search is a separate switch and is off by default. Only you can raise permissions; an agent cannot raise its own or anyone else's.
| Level | What it can do | Notes |
|---|---|---|
| Chat only | Chat and use team tools | Safest; the default for new agents |
| Read workspace | Read files in the work folder | Cannot edit files |
| Read and write workspace | Read and write the work folder and run commands inside a sandbox | Sandboxed. No network by default; add allowed domains to reach the internet (for example the one-click GitHub preset) |
| Full access | Run any command and access the whole computer | No sandbox. Use only when the sandbox tools are unavailable and you fully trust the task |
- Available levels differ by provider: Gemini CLI has no sandbox, so it has no read and write workspace level; API key agents are chat only.
- Work folder: the team shares UniverseBot/workspace in your home folder, and you can give individual agents a different one.
- Blocked by the sandbox by default: the app's own folder, ~/.ssh, ~/.config, ~/.gnupg and more.
- Allow web search: one switch per agent, off by default, and only you can turn it on, on screen. Once on, the agent can use its CLI's built-in search and page fetching, which is not limited by the allowed domains list.
- Agents can adjust another agent's permissions, but no higher than read workspace. Commander status, external MCP, the browser, the web switch and allowed domains can only be granted by you, on screen.
- On Linux, the read and write workspace level needs bubblewrap and socat installed first; macOS has a built-in sandbox.
What happens when something fails? Does it retry?
Errors are sorted first. Temporary problems such as overloads, server errors and dropped connections are retried automatically up to 2 times. Running out of quota, not being signed in, an unavailable model or an overly long chat are not retried; you get the reason straight away along with a button to fix it.
| Error type | Shown as | What happens |
|---|---|---|
| Temporary | Connection error | Retried automatically up to 2 times |
| Out of quota | Quota reached | Shows when it should be back; switches to the fallback model if you set one |
| Sign-in or key problem | Sign-in error | Asks you to check the model settings |
| Model unavailable | Model error | Offers a Change model button |
| Chat too long | Chat too long | Offers a Clear chat button |
| Anything else | Run error | Shows the reason |
What is a fallback model?
A fallback model is a second model you choose for an agent ahead of time. When the main model runs out of quota, the agent switches to the fallback so the work keeps going. It is off by default, and every switch leaves a note in the chat.
- It only switches when the quota runs out, not on other errors.
- The recent conversation goes along with the switch so the fallback can pick up the thread.
- When the main model is available again, the agent switches back and catches it up on what happened in between.
- ⚠️ A fallback model may be billed separately, and free models may train on your conversations. Keep that in mind when you set one.
Can avatars be animated?
Yes. Avatars can be animated GIF, WebP or APNG files (up to 5 MB). They only move while the agent is thinking, working or sending a message. If your system has reduced motion turned on, only the first frame is shown.
Everything else
- Image attachments: up to 10 images per message, 5 MB each (not supported by Gemini CLI agents).
- External MCP servers: attach your own MCP servers (stdio or HTTP) to Claude CLI agents by pasting their JSON. External MCP servers run outside the sandbox with the same permissions as you.
- Project notes: if an agent's work folder contains AGENTS.md or CLAUDE.md, they are read in automatically (up to 8 KB combined).
Coming soon
These are not available yet. We will update this page when they are.
- Use the desktop app remotely from your phone (over Tailscale): for now the desktop app only accepts connections from the same computer.
- LINE: talk to the coordinator on LINE.
- More command-line tools: such as opencode and hermes.