FAQ

​

Is this a chatbot with plugins?

No. You get a Linux machine that keeps running between conversations, and an agent that works on it — a shell, a filesystem, a browser, a package manager. A turn is not a request to a model with some tools bolted on; it is a program running on a computer that still has yesterday’s files. → How it fits together

​

Do I have to write code to use it?

No. Sign in, type in the chat, and the agent works on the machine you were given. Writing code is how you change who the agent is — its prompt, its tools, the programs it runs — and you can go a long way before you need to. → Quickstart

​

Does the machine stay running when I close the tab?

On Pro and above, yes: nothing stops it, and anything the agent left running keeps running. On the free plan it stops itself after 30 minutes with nobody talking to it, and starts again on your next message. Either way the disk survives. → Machine

​

What happens to files between conversations?

They stay. /home/daytona is the agent’s home and everything under it survives a stop, a restart and the end of a chat — files, checkouts, installed packages, a database a program wrote. The read-only root is the exception: /work, /data and /app do not exist. → Memory and what persists

​

Can I have more than one machine?

Free and Pro run one at a time, Max three, Max 20x five. A machine runs one harness at one ref, so a second harness of yours needs a machine of its own. A plan that shrinks does not delete anything; the extra machines simply stop starting. → Plans

​

Where does the agent’s prompt live, and how do I change it?

In the harness, as code: agent("Assistant", system_prompt=…) inside programs/splox/main.py. Ask your agent to edit that file, or edit it in the checkout at ~/harness yourself. There is no prompt box in settings, on purpose — the prompt is versioned with everything else that decides behavior. → Agent

​

When does an edit take effect? Do I have to publish?

An edit to the checkout on the machine lands on the next turn — the platform asks main.py who an agent is at the moment it needs to know. Publishing (git push origin main) makes a numbered version, which is what a new chat and any other machine on that harness pick up. So: try it in place, publish when it is right. → Publishing a version

​

What is a starter, and what happens when I copy one?

A working harness you copy to get your own. Copying creates a repository at the starter’s current commit and a machine to run it on; from then on it is yours and diverges however you edit it. The provenance is recorded, so you can merge the starter’s later changes if you want them. → Starters

​

Why does the agent write Python instead of calling tools normally?

Because a tool here is a Python function it imports, so one block of code can call four tools, loop over the results and print only the answer — instead of four round trips through the model with every intermediate result pasted back into the context. It is cheaper, faster and easier to read afterwards. → Tools

​

How do I connect Gmail, Linear or Notion?

On the Tools screen, as MCP servers. Pick one from the catalog and authorize it; the credential stays on the platform and is attached to the call, never written into the machine. → Tools · Connections

​

What is the Connections screen, then?

Model providers and environment secrets. Add your own OpenAI, Anthropic or Gemini key and it becomes an endpoint the composer can pick; add a secret and it is injected into the machine’s environment for your own code to read. Connecting a service is the Tools screen, not this one. → Connections

​

Which model does it use, and can I bring my own key?

The composer’s Harness decides means no selection at all — the harness picks, in code. Choose a model there and that chat uses it. A key you add on Connections becomes an endpoint you can select, and calls through it are billed by your provider rather than by us. → Model

​

Can I run a Telegram bot, a watcher, or something nightly?

Yes — that is a program in the harness, started once and left running. It survives the turn that started it, the end of the chat and the next day; it dies only when the machine does. Write it so that starting it twice is harmless. → Programs · Patterns

​

Can I drive an agent from my own code?

Yes, over the HTTP API with a Bearer key. Note that a run needs a machine_id (the public m_… id) as well as a harness, and that no published SDK sends that field today — so use the API directly for starting runs until they do. → Runs · Python SDK

​

Why did my turn stop with “You have used your 5-hour limit”?

A plan window filled. Nothing is lost — the messages, the tool calls and the files on the machine all stay — and the message names the moment the window frees up. Retrying sooner, a new chat or a second machine change nothing, because the windows belong to the account. → When a limit is reached

​

Does topping up raise the limit?

No. Money in the wallet does not change the cap; it pays for calls past a full window at list prices, which is what extra usage is. A bigger plan raises the cap, and what you have already spent is measured against the new one. → Extra usage

​

Who can see a chat I shared?

Anyone with the link, with no account and no sign-in — the token is the access control. They see the whole transcript, including tool calls, so read the turns before you send it. Make private kills the link at once, and sharing again mints a different one. → Security

​

Is my data used to train models?

The text of a turn goes to whichever model provider the run resolved to, and the app asks for explicit consent before the first time that happens. The privacy policy names the providers and the safeguards; it makes no statement about training in either direction, so do not assume one — ask support if it decides whether you can put a piece of work here. → Data and privacy

​

What happens if I delete a machine? Do I lose my chats?

The sandbox and its disk are destroyed; the harness is untouched and your chats are kept, though a chat whose machine is gone has nothing to run on. Deleting a harness takes the repository and every machine running it. Neither is reversible. → Machines

​

The agent says it wrote a file and I cannot find it.

Four usual causes: it used ~, which the starter’s file tools do not expand (/home/daytona/notes works); a sub-agent wrote it in its own worktree at /home/daytona/workspace-<name>; it tried to write outside the home directory, where the filesystem is read-only; or it ran on a different computer — another machine, or a local:/ssh: target. → Troubleshooting

​

Can two people share one agent?

Not as an account. There are no teams, seats or invitations: an account is one person, and what you can hand somebody else is a read-only share link to a conversation. Two people wanting their own agents want two accounts. → Chat

​

Is there a desktop app, and can the agent work on my own computer?

There is a desktop app for macOS, Windows and Linux; it is the same product in a window, plus a small daemon that lets a tool call address your own machine as local:<device_id>. Machines on your own hardware are not switched on yet. → Desktop app