Almost two years ago, we shared Mastra, our open-source TypeScript agent framework, on Show HN: https://news.ycombinator.com/item?id=43103073. It got a few hundred upvotes; a couple of weeks later our downloads had doubled, and we were off to the races.
Today we're sharing Mastra Factory, the open-source software factory we built to help maintain Mastra itself.
At Mastra, we now average nearly 4M downloads per week, and with Factory our team now merges hundreds of PRs a week across the mastra-ai GitHub org. Last week, Factory opened and merged 238 PRs, alongside 630 merged PRs from humans.
There's a lot of debate about what a software factory is. For us Factory is a configurable agentic kanban board that's @-taggable in Slack.
Under the hood, it uses Mastra’s durable agents for long-running sessions, our conversational memory system, a markdown based wiki with source citations to incorporate company knowledge, signals that subscribe to PR comments wired up into sessions, and a customizable cloud sandbox with our coding harness (working on supporting Claude Code, Codex etc).
When we started out, we modeled Factory around our existing workflow-oriented process (triage, plan, build, review, shipped). Three months later, we find ourselves mostly driving it through Slack vs a web UI.
We’ve seen more senior engineers kick off sessions in parallel (“triage the 20 oldest issues and let’s evaluate…”). We’re using bug and vuln scanners now, and we’ve closed our repo to outside contributors, asking folks to write detailed issue repros instead.
You can get started with `npm create factory`.
It’s fully open-source, so you can run it entirely on your infra (we also offer a hosted version). Comment here, we would love to know what you think!
Couple of questions:
1. Did you experiment with not having a separate design/plan phase? From my experience agents usually don’t really know what the plan is until they start editing files and executing tests (in other words, implementing), so the initial plan is usually not great and changing it mid-flight is overhead. I also find reviewing tested code to be much easier than reviewing plans/designs, which are often full of ungrounded assumptions.
2. Is this using your own custom coding agent harness over the raw LLM endpoints users provide? If so, did you consider having it integrate with existing harnesses (e.g. Claude Code, Antigravity, etc.)?
"I wont know until I look, and it'll take a few iterations to get it right"
https://factory.mastra.ai
https://github.com/mastra-ai/template-factory
but it all seems to come down to the same primitives around durable agents and workflows, memory, knowledge, sandboxes, filesystems.
the one exception is actually factory, it's a bit of a departure from what we've done (since it's application on top of mastra primitives vs the primitives themselves).
I've seen this pattern happening in other open-source repos. Do you think it's going to be the norm now?
also there are a few folks on our team who we got to know because they showed up in our repo and started contributing, and maybe that sort of serendipity won't happen in the future
I suspect sticking your public repo somewhere other than GitHub and sharing it in more organic programming communities (discord, forums, your hn bio etc) would further reduce the amount of low effort slop you get.
There is a [language](https://redbeardlab.github.io/dotFactory/) that let express complex workflows.
The workflow can be pretty much anything, eg:
1. Clone repo A and B + Start OpenCode + Feed a prompt into it
2. Pick up the commit generated by opencode for both repo, and run CI
Another workflow I created is: every day at 3AM, check that the documentation of the website is up to date, if yes exit, if not update the documentation. Pass the updated documentation to an adversarial LLM and if it pass, merge it - otherwise loop with the feedback from the adversarial one.
You have a concept of trigger, which can be either a webhook, or a github PR or a github issue, or just a cron expression. Or just someone manually clicking.
There is a concept of secrets and config, so the user can create their own secrets to start codex-cloud, or opencode, or making custom API call.
There is also the concept of manual gate, so you can wait for an human to actually click a button.
Finally, it supports running containers. Useful if you want to call an obscure API from your workflow, you don't have to wait for the platform owner to create a step for your API, but you can just run a container into it.
Open to input on what you have found is best, but in general a belief we have is to not be too dogmatic one way or the other and enable users to choose.
we will be posting a shorter demo/best practices video to our youtube soon as well!
Software factories are not a real architecture. They are just AI agents working on the same project. GitHub and similar platforms are already ideal for this. Requirements go in as issues, agents open PRs, someone reviews and merges them.
We used to call this software development.
The only "new" part is that now nobody complains at standups!
rumors of our demise have been greatly exaggerated...
at the end of the day it's open source and it's working well for us!
you can customize this a lot based on your needs. You can create skills for each column in a factory board, select different models, and steer the agents as they are running.
Depending on how you have it set up, your agent will stop for you to approve plans, answer questions, etc.
In our factory, low risk/low complexity items are more automated while higher risk/higher comlpexity items need more supervision.
Mastra Factory provides fully working preview environments in a sandbox so you can quickly test UX/UI changes in your app.
see https://yegge.ai/wyvern as an (imo) pathological case of this.
(not a criticism, just curious :))