Feedback wanted: an open-source agent team that builds and verifies ServiceNow apps
- Mark as New
- Bookmark
- Subscribe
- Mute
- Subscribe to RSS Feed
- Permalink
- Report Inappropriate Content
20m ago
Hi everyone. I've been working on a personal side project and would love feedback from other developers.
kiwistack is an open-source (MIT) plugin for Claude Code. It runs a ServiceNow build as a team of AI agents instead of one:
- recon reads the real instance;
- a domain agent pins down definitions, such as what "needs attention" means;
- a designer writes the spec;
- a builder is the only agent allowed to write;
- QA, security, two reviewers and a slop checker verify the work.
Nobody checks their own work, and a run isn't done until every number has been re-checked against the instance and every change is captured in the update set.
The ServiceNow side goes through a small guarded server:
- It refuses production instances.
- It writes only inside a kiwi/ update set.
- It asks before touching out-of-the-box records, data records or deletes.
It can build Service Operations Workspace pages from a short spec, build scoped-app logic through the SDK, and run a read-only audit of any scoped app, including Build Agent apps.
One lesson that might help others: kiwistack's own audit found its first big run left 135 of 325 changed records outside the update set. Most came from the SDK install. They move with the app, not the update set.
Source and install steps: github.com/Nivetha200111/kiwistack
I'd especially like to hear:
- where the page compiler falls short for your module;
- what guardrails you'd want before letting an agent loose on a shared dev instance;
- update set or SDK workflows I've got wrong.
It's a personal project, not affiliated with or endorsed by ServiceNow. I used AI to help build it.