PatchHive is a desktop-first multi-agent workspace for open-source contribution work. It gives each agent a clear specialty, keeps evidence attached to every decision, and leaves risky edits or maintainer communication behind explicit human approval gates.
Most multi-agent demos stop at chat bubbles. PatchHive is aimed at real contribution workflows:
- issue triage
- repository understanding
- review simulation
- patch planning
- test generation
- maintainer handoff
The product idea is simple: let multiple agents work in parallel, but keep the process structured, inspectable, and human-led.
This first version is a product prototype built with React, TypeScript, and Vite. It showcases:
- a mission-based workspace with reusable presets
- parallel agent lanes for planning, repo reading, review, patching, and testing
- timeline checkpoints and next-action prompts
- evidence cards tied to files, logs, and decisions
- human approval gates before risky handoffs
- a maintainer-facing outputs panel
The interface currently ships with three example missions:
PR RescueIssue IntakeRelease Brief
PatchHive is designed around five principles:
- specialization over generic agent chatter
- evidence before action
- human approval for meaningful risk
- maintainers need concise handoffs, not noisy transcripts
- workflows should feel good on desktop and remain usable on mobile
- React 19
- TypeScript
- Vite
- hand-authored CSS
npm install
npm run devFor a production build:
npm run build- add a real mission composer instead of static presets
- persist mission state locally
- attach repository files and issue links as live evidence
- support pluggable model providers
- add replayable activity logs and exportable handoff summaries
- explore a Tauri desktop shell
This repository used to contain a small lottery page. It has now been reset and rebuilt around a new product direction: a serious multi-agent collaboration tool for software contribution workflows.