Ship code from
your phone.
Say what you want changed. An agent clones your repo into a cloud sandbox, writes the code, runs the tests, and hands you a diff. You review it with your thumb and open a pull request — on the bus, in a queue, from bed.
Add tax to checkout total
shop-api · 3 files · +33 −2
Agent · Region comes from the existing cart.region field, so no schema change was needed.
The premise
The laptop was never the point. The keyboard was.
Coding agents removed the need to type most of the code. What is left is describing the work and judging the result — and describing work is something you can do faster by talking than by typing.
Every other tool in this category treats mobile as a smaller desktop: a responsive layout bolted onto an editor that assumes a trackpad. That is why they feel awful on a phone. The hard problem is not editing on a small screen. It is reviewing on one. We built the whole product around that.
1 in 3
developers outside the US and EU do most of their computing on a phone, not a laptop
~4×
faster to speak a task description than to thumb-type it
0
tools in this category designed review-first instead of editor-first
How it works
Three moves, none of which need a keyboard.
Say it
Hold the mic and describe the change in plain language. Pocket transcribes on-device where it can, and never makes you type a prompt on a phone keyboard.
"Add tax to the checkout total, use the region already on the cart"
It builds
Your repo is cloned into a fresh sandbox on Google Cloud. The agent reads the codebase, edits files, runs your tests, and fixes what it broke. Close the app — you get a notification when it lands.
Ran 84 tests · 2 failed · fixed · 84 passed
Review with your thumb
One hunk at a time, full width, big type. Swipe to move, tap to accept or reject, hold to ask why. Then open a pull request, or tell it what to change and send it back.
3 files · +24 −6 · pull request opened
Why Pocket
Six decisions the desktop tools would not make.
Cursor, Replit, and Codex all ship cloud agents. None of them were designed by someone whose only computer was a phone.
Review built for thumbs, not shrunk for them
Hunk-at-a-time paging, 44px targets, syntax highlighting tuned for sunlight, and inline agent rationale on every change. The review screen is the product, so it got the design budget.
Voice is the input, not a gimmick
Speech goes straight to the agent with your repo's vocabulary loaded, so it hears `useEffect` and `Stripe webhook` correctly instead of guessing. Talking is the only input method that is genuinely faster on a phone.
Real repos, real stacks, real root
Not a walled runtime. Your actual GitHub repo, your actual Dockerfile, your actual test suite, with full shell access. If it runs on Linux it runs here.
Bring your own key, or your own machine
Plug in an Anthropic or OpenAI key and pay nothing for tokens. Attach your own GCP box, a Vast.ai instance, or the desktop under your bed, and pay nothing for compute either. We would rather keep you than trap you.
Priced for where phones are the computer
₹350 in India, R$30 in Brazil, Rp 80,000 in Indonesia — paid with UPI, Pix, and GoPay, not just a card that many developers do not have. The margin math works because compute is cheap; see our pricing page, we publish it.
Async by default
Dispatch and leave. Runs continue with the app closed and the screen off, and push a notification when they need you or finish. Long-running agents are the normal case on mobile, not an edge case.
Sandboxes
Your stack, already installed.
The default image ships with the common runtimes and package managers warm. Need something else? Point a workspace at your own Dockerfile.
Pricing
You pay for the sandbox. Tokens are close to cost.
Compute is the part we can buy well and run cheaply, so that is what we sell. Model tokens carry a small markup, and you can remove it entirely by bringing your own key.
Free
Enough to find out whether you'll actually ship from your phone.
Free forever
- 30 minutes of agent time per month
- 1 connected repo
- Bring your own model key for unlimited tokens
- Full diff review and PR creation
Starter
For the side project you keep thinking about on the train.
Billed monthly, cancel anytime
- 6 hours of agent time per month
- $2.50 of model credits included
- 3 connected repos, 2 runs at once
- 2 vCPU / 4 GB sandboxes
- Unlimited tokens with your own API key
Pro
Your primary development environment, which happens to be a phone.
Billed monthly, cancel anytime
- 25 hours of agent time per month
- $6.50 of model credits included
- Live preview URLs for running apps
- 4 vCPU / 8 GB sandboxes, 5 runs at once
- Frontier models: Sonnet 4.5, GPT-5, Gemini 2.5 Pro
Plan allotments reset each billing period and do not roll over. Separately purchased top-up credits never expire — we do not run an expiring-credit business.
Questions
Reasonable doubts.
Is this just an editor on a small screen?
No, and that distinction is the whole product. There is a code editor for small fixes, but the primary surface is a review console: you dispatch work by voice, the agent runs in the cloud, and you approve or reject changes hunk by hunk. Editing is the fallback, not the default.
What happens when I close the app or lose signal?
The run keeps going — it executes on Google Cloud, not on your device. Actions you take offline are queued and reconciled when you reconnect, and you get a push notification the moment a run finishes or needs a decision.
Do you train on my code?
No. Repos are cloned into a single-tenant sandbox that is destroyed when the run ends, and nothing is retained beyond the diff and logs you can see in the app. Bring your own key and the model provider never sees us in the middle at all.
Can I use my own Anthropic or OpenAI key?
Yes, on every plan including free. Your key is encrypted at rest, used only for your runs, and tokens billed to it do not touch your Pocket credits. You still consume sandbox time, which is what your plan actually pays for.
What if I need more compute than the plan includes?
Attach your own. Pocket can dispatch runs to a GCP instance, an AWS box, a Vast.ai GPU, or a machine at home over a tunnel. Those runs bill nothing against your agent-time allotment.
Why is it so much cheaper in India and Brazil?
Because a flat global price in dollars is a decision to not serve those markets. Compute costs us a few cents per run, so regional pricing still clears costs comfortably — we publish the unit economics rather than asking you to take our word for it.
Your next commit doesn't need a desk.
Connect a repo, say what you want, and see whether you actually ship from a phone. Thirty minutes of agent time, no card.