I had access to five ChatGPT Pro plans and three Claude Pro plans through referral offers. Ten days later, I had spent more time switching accounts and remembering which project I had left open than using the coding allowance. I wanted one place to describe work, let an available tool take the next step, and review what it actually produced.
That became Runquay: an MIT-licensed, local work queue for AI coding tools. It launches the official Codex, Claude Code, Gemini CLI, or Antigravity CLI on your computer. Runquay itself doesn't need a separate model API key or a hosted account. Your selected provider still handles the AI request under its own terms and billing.
The actual Runquay Create screen. The visible history is fictional demo data, so no private projects appear in the screenshot.
What happens after you type a request?
You can ask for a new project in plain language or choose an existing folder. Runquay saves the request, runs one bounded step at a time, and keeps the files and a checkpoint in that project's workspace. You can pause, inspect the result, ask for a change, or continue later. It can also find local sessions created in the original coding tools, so you can browse your work in one place while continuing to use those tools directly.
The distinction that matters to me is where continuity lives. If the next step runs on the same Codex or Claude profile, Runquay can resume that tool's native conversation. If work moves to another account or tool, it carries on from project files and a saved handoff. It does not transplant one account's private conversation or credentials into another account.
Here's an older, real Codex walkthrough: a request for a small fictional café website finished, and Runquay showed how to open and check the generated project.
A real completed Codex run, captured in an earlier Runquay interface. The café and its content are fictional.
Account switching has to respect the limits
For Codex, Runquay checks subscription quota before work and maintains a reserve. When one connected profile reaches that reserve, it can select another eligible profile. If none qualifies, the task waits for allowance to refresh. Paid credits and earned resets are separate decisions; Runquay will not silently use them.
Claude Code is different. Runquay cannot reliably check its subscription-versus-paid billing before a run. It asks for approval before every Claude invocation. If Claude Code explicitly returns a usage-limit error, Runquay can select another connected Claude profile, then asks again before that profile runs. An ambiguous error is not treated as permission to switch or retry.
You can connect as many profiles as you need, but Runquay uses one worker at a time. Extra profiles do not increase a provider's allowance, and you should use only accounts you control under the provider's rules. Gemini CLI, Antigravity CLI, and custom tools can join the queue; their billing remains unverified, so each run needs approval, and their accounts do not yet rotate automatically on errors.
The useful part beyond the queue
Work across projects can teach you things worth keeping. Runquay's Collective Learnings brings matching local work summaries together, then lets a connected AI analyze the selected context when you ask it to. You can review sources, remove a summary you don't want to send, save useful lessons, and attach a lesson to later work. The topic display starts with local matches; it is not claiming that an AI has already drawn conclusions just because the page opened.
The actual Learn screen with fictional summaries. Analysis begins only after the user chooses it.
Getting started, and what's still unverified
The latest release has guided installers for Windows and macOS/Linux. Sign in through each vendor's official flow, choose a folder, and describe the first task. The beginner guide walks through the screens without assuming you know how to code.
Real Codex and Claude project runs have been tested on Windows. Automated checks run on Windows, macOS, and Linux, but native vendor runs on macOS and Linux still need hands-on testing. Two-account switching has been exercised in an isolated sandbox with CLI stand-ins; I have not yet observed a live switch between two real vendor accounts. The verification record separates those checks from live provider tests.
The source, setup instructions, and issue tracker are at github.com/dibyapp/runquay. If you try it, I'd especially like to hear where setup, approvals, or handoffs are confusing or fail in practice. Concrete bug reports and contributions are welcome.
Disclosure: AI helped draft this article. Product claims were checked against the linked public source and verification notes.
Top comments (0)