New in 0.4.22
AI accounts and limits
Account lists predate 0.4.22. This release fixes Claude limit switching and adds separate CLI sign-ins. Control names in this guide match the English interface.
Set the order
Add accounts you already use
Open Settings → AI → AI accounts. Under Add a sign-in this computer already has, choose a signed-in CLI on this machine or a saved API key and press Add; the credential stays where it is. Enable the accounts you want, then order them with the ↑ and ↓ buttons beside each row, whose tooltips read Try this one earlier and Try this one later.
Enable switching
Switch automatically when a limit is reached is on by default. With it enabled, the AI panel tries the next ready account in your order after a usage limit or sign-in failure. The same question and bounded conversation history are sent again; the next account does not inherit the previous account’s private memory. A switch notice names the accounts for five minutes.
Sign in with the CLI’s own login
Sign in first, add more later
Settings → AI starts with Sign in. Press Sign in with Claude, Sign in with ChatGPT (Codex) or Sign in with Google (Gemini). A tool that is not installed appears under Not on this machine yet, with the line that installs it. Name the account, press Open a shell and sign in, finish the CLI’s own sign-in and press I have signed in. Gemini signs you in inside its own session: when it says you are signed in, type /quit. Termiyo creates a private folder and environment for each account. An API key is optional, under Advanced: use an API key instead.
Add a second account of the same tool
Once the list holds an account of a tool, its button under Add an account reads Add another Claude account, Add another ChatGPT account or Add another Gemini account. Each press signs one more account in, in a folder of its own. Use a different account each time: one that is already in the list is refused, and the message names its row. Each row shows the address the tool reports; press Rename to give it your own name, and use ↑ and ↓ to choose which account answers first.
Know the restrictions
Harness and Work together mark a Codex account unverified, because separate files do not prove separate system-keychain logins; this screen carries no such mark. ChatGPT CLI, Ollama and llm have no separate sign-in flow here; adding the same login twice does not double its allowance.
When switching does not help
Errors and exhausted accounts
A temporary 429 that is not a usage limit fails the question instead of switching. Other request errors also stop that attempt. With automatic switching off, a refusal stops it too. If every account refuses, the panel lists the accounts it tried. When one of them was refused for being signed out, it asks you to sign in again to that account, and adds the first reported reset time if there is one. Otherwise it gives a return time only when one was reported, and says it will try again on your next question when none was.
Resting is not a reset promise
For a usage limit, Termiyo uses the provider’s reported reset time. Without one, its retry delays rise through 1 hour, 5 hours, 24 hours and 7 days after repeated refusals. These are retry estimates, not guaranteed reset times. A signed-out account needs a new sign-in; waiting does not sign it back in. Ready accounts are tried first; resting accounts can still be tried as a last resort.
Read the figures correctly
Last reported, not live
Claude can report percentages used and reset times for its usage windows, shown with the reading’s age. Codex does not report those figures here. The status-bar chip names the account that answered last, not a promise of which will answer next. These readings are not live quota monitoring and do not balance work by quota.
Keep the three flows separate
The AI panel resends conversation text. A switched Harness lane starts fresh, and a reassigned Work together task restarts from its brief without partial output. None of these creates shared memory between accounts.