Agents do the work.
You decide.
Origin ranks every decision your agents wait on, in one inbox.
- Blocking now
- Work is stopped until you act.
- Needs you today
- Waiting on your decision, though nothing is stopped yet.
- Watching
- No action needed. Here so nothing is invisible.
From a label to a merge
One issue, followed through the factory.
An issue arrives
Label an issue on GitHub. Triage drafts the task, and asks you only what it cannot decide.
Agents build and review
Claude Code implements, the build checks it, Codex reviews. Findings send it back to the implementer.
You decide where it matters
Approvals, plans and blockers come to you, ranked by how much work your answer releases.
Every run is on the record
Cost, tokens and model for every stage, and how much of it bought nothing.
What it runs, and where
- Which agents does it run?
- Claude Code and Codex, on your subscriptions. opencode is switched off for now.
- Where does the code run?
- A fresh boxd machine for each stage, holding no credentials.
- Where do pipelines live?
- In your repository, under
.origin/, read at a pinned commit. - Is it ready?
- A prototype. Every screen runs on a real schema; the demo data is fiction.
Three stages, two vendors, one file
apiVersion: origin.dev/v1alpha1
kind: Pipeline
metadata:
name: builders
spec:
title: Builders
stages:
- key: implement
kind: implement
harness: claude-code
model: claude-sonnet-5
- key: checks
kind: test
shell: bun run build
- key: review
kind: review
harness: codex
edges:
- { from: implement, to: checks, on: success }
- { from: checks, to: review, on: success }
- { from: checks, to: implement, on: findings }
- { from: review, to: implement, on: findings }