1,000 Users In: What Queued Is, and Why It Runs on Claude — Not OpenAI

We crossed 1,000 users this week. No launch video, no ad spend — just developers who were tired of losing good ideas to the gap between *'I should fix that'* and *'I'm actually at my keyboard.'* It's a good moment to say plainly what Queued is, how it works, and the one thing people ask about most: why it runs on Claude instead of OpenAI.

What Queued is, in one paragraph

Queued is a mobile-first, autonomous AI coding app. You describe a coding task in plain English from your phone — *'add a rate limit to the login endpoint'*, *'fix the timezone bug on the reports page'* — and an AI agent clarifies the scope, writes the code against your actual repo, and opens a pull request on GitHub for you to review. You're involved at exactly two moments: describing the task, and approving the result. Everything in between runs without you. No editor, no laptop, no watching a cursor blink.

How a task moves through Queued

1. 1Capture — You write a task in plain English, from wherever you are. A sentence is usually enough.
2. 2Clarify — Before touching code, the agent asks a couple of sharp questions about anything ambiguous — scope, patterns, edge cases — so it doesn't guess wrong.
3. 3Queue — The task joins your queue and runs in order. Stack a few, reorder them, put your phone away.
4. 4Execute — The agent indexes your repo, writes the change to fit your existing structure, and checks its own work.
5. 5Review — You get a pull request on GitHub. Read the diff, request changes, or merge. You stay in control of what lands.

Why Queued runs on Claude, not OpenAI

This is the question we get most, so here's the honest version. Autonomous coding is a different problem from interactive coding. An in-editor assistant only has to be right for the next few lines while you watch. An unsupervised agent has to stay coherent across dozens of steps — reading the codebase, planning, writing, checking itself — with nobody there to catch it drifting.

We built and tested against both OpenAI's models (the GPT and Codex line) and Anthropic's Claude. Both are genuinely strong. For our specific loop — long, unattended, judgment-heavy — Claude was the more reliable partner. So Queued runs on Claude today. The differences that mattered:

• →Interactive vs. autonomous. OpenAI's tools shine when you're in the loop, prompting live. Queued is the opposite — nobody's watching, so consistency over a long run beats peak cleverness on a single prompt.
• →Reading your codebase. Claude has been reliably good at picking up the conventions already in a repo and matching them, instead of imposing generic patterns of its own.
• →Knowing when to ask. The clarify step needs a model that flags ambiguity rather than barrelling through it. That restraint is worth more than raw speed in an unattended job.

We're not religious about it. Queued is model-agnostic by design, and the moment a different frontier model — from OpenAI, Anthropic, or anyone else — is clearly better for unattended, PR-first coding, we'll move. You shouldn't have to care which model is under the hood. You should just get a good pull request.

What 1,000 users actually queue

The tasks people trust to an agent turned out to be narrower and more useful than we guessed:

• →Small fixes that never make the sprint — the off-by-one, the missing loading state, the typo in an error message.
• →Well-scoped features — *'let users reset their password from the login screen'*, described in a sentence.
• →Chores — bumping a dependency, adding a test, wiring a new endpoint into an existing pattern.
• →The 11pm idea — the thing you'd otherwise write on a sticky note and lose. Queue it from bed, review the PR over coffee.

What 1,000 users changed about the app

A few things we got wrong and fixed because people told us:

• →The clarify step earns its keep. Early on we let tasks run with less up-front clarification and the pull requests were worse. Two sharp questions before the agent starts is the single biggest lever on output quality.
• →Big tasks need breaking down, not rejecting. A task too large to ship as one clean PR used to just fail. Now Queued offers to split it into smaller queued tasks — because 'this is too big' is a plan, not an error.
• →Order matters. People wanted to arrange work before it ran, so tasks that haven't started yet are now editable and drag-to-reorder in the queue.

Where we're headed

The next 1,000 users get a sharper version of the same loop: better clarifying questions, faster execution, and the same non-negotiable rule — every change arrives as a pull request you review. If you've been meaning to try it, Queued is a free download on the App Store and Google Play, and every paid plan starts with a 3-day free trial.

Frequently asked questions