hobai is live
Describe a clear coding task, and hobai opens the pull request. Here's how it works end to end, and what it can and can't do yet.
A lot of coding tasks are clear before I even start. Add an endpoint. Fix a validation bug. Write the missing tests. I know exactly what should change, and it still takes a couple of hours of my attention.
So I built hobai. You describe the change; it opens the pull request. An agent does the work in an isolated cloud sandbox while you watch, runs your checks, and hands you a pull request to review.
How it works
- Connect GitHub. Install the hobai GitHub App on just the repositories you choose.
- Set up once. hobai works out how to install and check your repository: package manager, lint, type-check, tests, build. If it can't decide, it asks you a few multiple-choice questions. Every step it keeps has actually run in a fresh sandbox.
- Describe the task. In plain words, the way you'd write it for a teammate.
- Watch it work. Every file it reads, every edit, every command and its output, live.
- Review the pull request. Your checks run before and after the change, and the results go into the PR with a summary and the changed files.
A task can be as small as this:
Add a GET /health endpoint that returns 200 with the database status,
and 503 when the database is down. Add tests.hobai sets up the repository, runs the existing checks, explores the code, writes the endpoint and its tests, runs them, and opens the pull request.
Inside a run
Every run goes through the same six stages. Click through them:
Results you can trust
I wanted an agent that doesn't pretend. How a run ends decides what you get:
| When | You get |
|---|---|
| The checks pass | A ready pull request |
| The checks fail | A draft pull request, so nothing unfinished looks finished |
| Nothing needed changing | No pull request, and an explanation of why |
| The run hits its limit | The pull request says so |
And the rules I'd want from any tool that touches my code:
- Your code runs only in sandboxes. Every run gets its own, isolated and short-lived.
- Never your main branch. Every change arrives on its own
hobai/branch, as a pull request you review. - The agent never holds a token that can write. hobai makes the commit through GitHub's API from outside the sandbox.
- Access you choose, per repository. The GitHub App only sees the repositories you pick.
- You can cancel any run, at any time.
What it supports today
| Area | Supported |
|---|---|
| Code host | GitHub, through the hobai GitHub App |
| Languages | JavaScript, TypeScript |
| Package managers | npm, pnpm, yarn |
| Repository layout | One app, or a separate frontend/ and backend/ |
It works best on well-defined tasks a developer would finish in one to a few hours: a feature or an endpoint, a bug fix, missing tests, a refactor, a dependency update and its fallout.
It's free to start: 10 runs a month, no card needed. Setting up a repository doesn't use a run. See pricing for the details.
What I'd like to hear
If you try it, I'd love to know:
- What task did you give it, and would you merge the pull request?
- Where did setup get stuck, or ask the wrong question?
- What would you need before trusting it on your team's repository?
This devlog is where I'll write up what I learn, what breaks, and what changes because of it.