Skip to main content
This page starts from what you see in Alpha, such as a red message, a blank pane or a button that does nothing. Each entry gives the likely causes and the fix, and many include a message you can send to the agent as it is. Feature pages keep their own full tables of messages, linked from each section. For the numbers behind a limit, see Limits.
The agent reads the same logs and errors you do, and it can usually fix its own work. When an entry offers a message, send it in the session that has the problem.

​
Can’t sign in, or Alpha won’t load

Sign-in errors appear in red on the sign-in screen. Every one of them is listed on Your account and settings.
Causes: Your Alpha only lets in approved email addresses, and yours isn’t on the list. This way of signing in has no waitlist to join.Fix: Ask an admin to add your address, or sign in with an address that’s already approved.
Causes: You signed in with Google or Microsoft from an address that isn’t approved yet. Alpha has made your account and holds it until an admin approves it. The team already knows you’re waiting.Fix: Once you’re approved, press Check again or refresh the page. “Still waiting for approval. Try again in a little while.” means you’re not approved yet. “Couldn’t reach Alpha. Check your connection and try again.” means the check itself failed. If you signed in with the wrong account, click Sign out under the card. See If you’re on the waitlist.
Causes: There were too many sign-in attempts for this email from your network in a short time. Alpha refuses more until the window passes, to protect the account.Fix: Wait a few minutes, and check the password before the next try. If you see “this account signs in with Google or Microsoft — use the buttons above”, the account has no password. Use the button instead.
Causes: Alpha’s server didn’t answer when the page loaded, and this browser had no remembered sign-in to open with. The server may be restarting, or your connection dropped.Fix: Alpha keeps retrying by itself, waiting a little longer between tries, up to 15 seconds apart. Press Retry to try at once, or check your connection. The red line under the message is the error your browser got. Pass it to your admin if the page stays like this.
Causes: Part of the page hit an error while drawing. “Alpha couldn’t load.” means the page’s outer frame failed, not only one view.Fix: Press Try again. If the message comes back, reload the page. Your work is safe either way. Turns run on Alpha’s servers, not in the page, and unsent text in the composer is kept. If one session or file brings the error back every time, tell your admin which one.
Causes: Alpha was updated while the tab was open. When the tab asked for a part of the old version that no longer exists, Alpha reloaded it once to fetch the new one.Fix: Nothing. The session and its open tabs come back after the reload. If the tab stays on “This view couldn’t be shown.” instead, reload it yourself.

​
The machine won’t wake, or the chat stops updating

A session’s machine sleeps once you’ve been away for about 30 minutes and nothing is running, and wakes when you come back. Alpha never swaps a sleeping machine for a blank one behind your back, so every message below means nothing was discarded. How sleep works is on Sleep, wake and reset.
Causes: The machine is waking. Tabs that read from it, such as files, Preview, Shell and Database, wait behind placeholder bars until it’s up.Fix: Wait a few seconds. The tabs fill in by themselves, with nothing to retry. The Shell can take a few seconds longer than the rest.
Causes: The saved machine didn’t start. The reason is in the brackets. The saved copy is kept.Fix: Send your message again in a minute. If it keeps failing after several tries, ask an admin to reset the machine (Admin → Sessions → Reset the machine), which gives the session a fresh one. Read Reset the session’s machine first, because a reset deletes every file on the machine.
Causes: The server that holds the machine is briefly out of reach.Fix: Your text goes back into the composer. Send it again in a moment. Don’t reset for this one. The machine is fine, only out of reach for now.
Causes: Something else woke the machine or put it to sleep at the same instant as your request.Fix: Send again.
Causes: Alpha’s server ended this tab’s live updates for the session, for example while restarting. A brief drop in your own connection reconnects by itself and doesn’t show this.Fix: Reload the page. The transcript replays and you land on whatever the agent is doing now. The turn kept running the whole time.
Causes: You’re viewing someone else’s session and its machine is asleep. People a session is shared with never wake its machine. The sidebar’s hover card reads “Machine asleep · wakes when its owner is back”.Fix: The transcript stays readable. Files and the Preview come back once the owner returns to Alpha. See Share a session.

​
The Preview shows no app

The frame, or a preview address opened in its own tab, says why there’s no app. The full list is on Preview your app.
Causes: The app’s server only listens inside the machine. The Preview reaches it from outside, so the server must listen on 0.0.0.0.Fix: Ask the agent to rebind it.
Causes: The app’s server crashed or was stopped, or it never started on that port. After a wake, Alpha gives the server a few seconds to come back before it says so.Fix: Click Logs in the Preview toolbar to see why it stopped, then ask the agent to run it again.
Causes: Every port that’s open serves an API, a file listing or a slide deck, not a web page.Fix: Pick a port from the port menu in the toolbar, or ask the agent to start the app’s front end. A deck opens in its own viewer from the file tree. See Slides, drawings and diagrams.
Causes: The dev server, Vite for example, only answers addresses on its allow-list, and the preview’s address isn’t on it.Fix: Ask the agent to allow the preview’s host.
Causes: “Sandbox asleep” means the address was opened while the owner is away from Alpha. Preview addresses never wake a machine. “Sandbox didn’t wake” means the machine couldn’t be brought back. Nothing was discarded.Fix: Open the session in Alpha. “Sandbox asleep” reloads itself once the machine is back. For “Sandbox didn’t wake”, Alpha says what went wrong when it tries to wake the machine. Follow the machine entries above.
Causes: Google’s sign-in page, like many outside pages, refuses to be shown inside another page. So the Preview shows “Continue on accounts.google.com” and opens it in a window of its own. Inside the frame, the browser can also drop your app’s sign-in cookies, so the app keeps forgetting you.Fix: Finish in the window that opened. The app picks up in the frame when you’re done. If sign-in keeps looping, click Open in new tab in the toolbar, where cookies work as in any tab. “This step is no longer open” means the window was closed or reused. Go back to the app and start again. See Sign in through another site.

​
A turn fails or a command is refused

A failed turn keeps the workspace and the transcript up to the failure, and your next message carries on from there. More messages are on Follow the agent’s work.
Causes: The agent stopped with an error. The card says what failed.Fix: It depends on the card:
Causes: A Claude Code session couldn’t use the credential it runs on. The rest of the message says which one.Fix:How to sign in is on Usage, costs and your own plan.
Causes: The ChatGPT account is on the free plan. Codex needs a paid plan, and the sign-in dialog warns “This account is on the free plan. Codex needs a paid ChatGPT plan, so turns will likely be rejected”.Fix: Click Sign out and use the server key in the dialog to go back to Alpha’s key. If you upgraded a moment ago, click Re-check. See Run Codex sessions on your ChatGPT plan.
Causes: Some commands, such as /fork, /review, /init and /delete, need the agent to be idle. The banner names the command you typed.Fix: Wait for the turn to end, or stop it with Stop or Esc, then run the command again. See Queue, steer or stop.
Causes: You pressed Stop with messages waiting, a queued turn failed, the server restarted, or the next message couldn’t start. The reason follows the dash.Fix: Deal with the reason, then click Resume. Send now sends one message and leaves the rest paused. See Why did the queue pause?
Causes: The message is longer than 100,000 characters.Fix: Attach the long text as a file and refer to it in a short message. See The composer.
Causes: Voice input isn’t turned on for your Alpha, or your browser can’t record for it. Once it’s running, the browser may have blocked the microphone, the connection may have dropped, or the dictation reached 2 minutes.Fix: For “Microphone access was blocked — allow it in the browser to dictate.”, allow the microphone for Alpha in your browser’s site settings. For “Voice input disconnected — press the mic to try again.”, click the microphone again. A dictation that reached its limit stops by itself, so start another. See Dictate.

​
Uploads, files or the database fail

Causes: The upload was interrupted, or the file is larger than 256 MB per file (default). Hover the tile for the reason. If you sent anyway, you see “An attachment failed to upload — the message wasn’t sent.” and your text goes back into the composer.Fix: Click the tile’s X, then attach the file again. For a file over the limit, upload a compressed copy, or let the agent fetch it from a link.
Causes: You opened another session while a file was still uploading.Fix: Go back to the session and send again. Your text is kept in its draft.
Causes: A single file can be up to 32 MB, and a folder up to 500 MB. Dependency folders such as node_modules are always left out of a folder download and don’t count.Fix: Ask the agent to split or compress what you need.
Causes: The workspace belongs to the agent, so every file viewer is read-only.Fix: Ask the agent for the change, or make it yourself in the Shell. See Why can’t I edit files here?
Causes: The session’s database hasn’t been switched on. Every machine has one, but it stays off until it’s needed.Fix: Open + → Database, or ask the agent to add a database to your app. See Your session’s database.
Causes: The database server has stopped, or is slower to start than the machine after a wake. The server’s error shows under the title. “PostgreSQL isn’t installed in this sandbox” means the machine was created before sessions had a database.Fix: Click Start server. For a machine without PostgreSQL, click Install PostgreSQL 16. It reads “Installing… about a minute” while it works.
Causes: You picked a published app’s production database in the Database tab, and Alpha couldn’t connect to it. The error shows under the title.Fix: Click Try again. If it persists, check the app on its Publish tab and see Production database and sign-in.

​
A publish fails

A failed version never takes traffic. If an earlier version was serving, it keeps serving, so your live app stays up. The Publish tab marks the step that failed and shows a hint under the error, and the agent reads the same log. These three come up most. Every hint is explained on Monitor and troubleshoot production.
Causes: The app didn’t start serving within 4 minutes. It may listen on localhost or a fixed port instead of the PORT it’s given, or its start command runs a development server. Or the app needs its tables when it starts, and production’s database is still empty.Fix: Bind 0.0.0.0 and read PORT. If the app needs tables at start, run migrations in the start command or push the session’s schema. See Production database and sign-in.
Causes: The lockfile is out of step with the manifest, or a dependency doesn’t exist under that name or version.Fix: Regenerate the lockfile in the session, then republish.
Causes: Your framework only answers hostnames on an allow-list, and the app’s .run.app address isn’t on it. Every visitor gets an error page.Fix: Add .run.app to the allow-list beside the entries already there, so the Preview keeps working, then republish. Examples for Vite, Django, Rails and FastAPI are on The service is up, but refuses its own address.

​
GitHub, an app or an MCP server won’t connect

Causes: The repository’s name is misspelled, or the GitHub account you connected hasn’t granted Alpha that repository.Fix: Check the spelling. Then click Manage … on GitHub beside the error, grant the repository, and add it again. See If a repository won’t attach.
Causes: The repository belongs to an account or organization you haven’t connected.Fix: Click Connect … beside the error, pick that account on GitHub and grant the repository. See Add organizations and more accounts.
Causes: Both your branch and GitHub’s copy of it have new commits. Git’s own words show in a red strip under the repository name.Fix: Ask the agent to merge or rebase, then push.
Other refusals are on Branches, commits and pull requests.
Causes: The app’s sign-in page opened in a small window that may be behind this one. If your browser blocked the window, Alpha opens the app’s page in this tab instead and brings you back when you finish.Fix: Look for the sign-in window behind Alpha’s, and finish there within 10 minutes. To keep sign-ins in their own window, allow pop-ups for Alpha in your browser.
Causes: You cancelled on the app’s sign-in page, or the sign-in took longer than 10 minutes. For an MCP server, the line reads “The MCP server’s sign-in did not complete. Try connecting again.”Fix: Close the sign-in window and connect again from Alpha. For GitHub, “Nothing was connected — the GitHub sign-in was cancelled.” means the same, so click Connect GitHub again. More fixes are on Connect an app.
Causes: The agent used a remote MCP server whose sign-in has expired and couldn’t be renewed.Fix: Click Integrations in the sidebar, open the server’s card on the MCPs tab and click Sign in again. Then ask the agent to carry on. See MCP servers.
Causes: The value hasn’t been added as a secret, or the published app wasn’t given it. When the agent needs a key with no ready connector, it posts a card reading “The agent needs” and the variable’s name. A published app only gets the secrets you pick for it, and a name you haven’t added is refused with ”… is not one of your secrets”.Fix: On the agent’s card, click Add and the variable’s name, paste the value and save. Then click Tell the agent to continue. For a published app, add the secret first, then pick it under Environment. Never paste the key into the composer. See Secrets.
If you don’t see the microphone, Publish, the Claude sign-in or Continue with Google, it isn’t turned on for your Alpha. The same goes for “GitHub isn’t set up on this server yet — ask whoever runs it.” Ask your admin.

​
Frequently asked questions

The agent keeps working. Turns run on Alpha’s servers, not in your browser, so closing the tab, the window or the laptop doesn’t stop them. A turn can run for up to 48 hours (default). Once no Alpha tab of yours has been on screen for about 30 minutes and nothing is running, the session’s machine goes to sleep with everything kept. When you come back, the transcript replays and you land on whatever the agent is doing now. See Sleep, wake and reset.
Only you, until you share it. A session is private by default. You can open it to anyone with its link, or to specific Alpha accounts. Viewers get a live, read-only view of the transcript, files and Preview, with your secrets masked. They can’t send messages or change anything. Admins of your Alpha can view every session. See Share a session.
Yes. Sleep saves a full copy of the machine’s memory and disk: files, installed packages, the session’s database and running processes. The machine wakes exactly as you left it. Only a reset or a delete removes files. See What survives what.
Not in the file viewers. They’re read-only, because the workspace belongs to the agent. You can change files in the Shell, which runs as the same account as the agent, or point the agent at exactly what to change with Select and Draw.
Yes. Deleting a session also deletes every app published from it, along with its address and its production database. What you pushed to GitHub stays. To keep an app online, keep its session, and move the session into a group if you want it out of the way. See Fork or delete a session.
Not yet. Every published app gets a run.app address. Where your Alpha has a domain of its own, the app also gets an address on it, the same for every app. You can’t attach a domain you own. See Where does my app live?
No. Alpha has no trash or archive, and Delete in the sidebar doesn’t ask first. A reset can’t be undone either.
Deleting a session removes its machine, its files, its published apps and their production data for everyone. Resetting the machine deletes every file on it, including unpushed work. Push to GitHub, or export the database, first.
Yes, as secrets, encrypted at rest and tied to your account. The machine gets them as environment variables, so anything the agent runs can read them, including the app it builds. Values of 8 or more characters are replaced with [redacted] wherever they appear word for word in the transcript. Email addresses, plain web addresses and everyday words such as production are left as they are, because they name things rather than grant access. A published app only gets the secrets you pick for it. A key typed into the composer lands in the transcript, so add keys under Secrets instead.
The ones in the Model popover, opened from the model chip in the composer. The list comes from the agent itself, so it differs between Alphas and changes when your Alpha updates. The row marked default is what a session runs when you haven’t picked one. With Claude Code, the list follows the Claude account the session runs on. See Choose the agent and model.
If you’re not sure, use the one marked default. Codex and Claude Code share the same machine, workspace, Preview, database and publishing. They differ in how you talk to them, such as $name or /name for a skill. You pick the agent on the home screen, and a session keeps it for life, so try the other one in your next session. See Codex or Claude Code?

​
Next steps

Sleep, wake and reset

What a reset keeps and what it deletes, before you reach for it.

Monitor and troubleshoot production

Every publish hint, plus the live logs and metrics of a published app.

Limits

The sizes, counts and times behind “too large” and “at most” messages.

Inside a session's machine

Ports, folders and tools on the machine, for when the agent needs a nudge.