Broccoli: Cloud-based AI coding agent automates pull requests end-to-end

The rise of AI-assisted coding has transformed how teams handle small tasks. Yet many developers still juggle multiple local environments, context switching between feature requests, refactors, and bug fixes. A new open-source project called Broccoli aims to change that by running coding tasks in isolated cloud sandboxes and automatically opening pull requests for human review.

From local chaos to cloud clarity

Broccoli was born out of a common frustration: managing multiple AI coding sessions across different repositories and tasks. The team behind it, which operates a voice data company, found themselves frequently switching between tasks, juggling worktrees, and keeping laptops running just to ensure continuous execution. Each task required its own environment, context, and monitoring—adding unnecessary complexity to workflows.

To solve this, the team built Broccoli as a centralized cloud-based harness. Instead of running tasks locally, Broccoli takes coding requests directly from Linear tickets, checks out the relevant repository, applies the ticket’s context, implements the solution, runs tests, and opens a pull request for review. This approach eliminates the need for developers to manage multiple local sessions, reducing context switching and enabling continuous progress even when laptops are offline.

Over the past month, the results have been compelling. According to the team, 100% of pull requests generated by non-developers are now shipped via Broccoli, offering a safer and more efficient route than manual execution. For developers on the team, about 60% of smaller tasks are now handled by the tool, while more complex features requiring extensive back-and-forth design are still managed locally using Codex or Claude Code.

How Broccoli works under the hood

The technical architecture of Broccoli is designed for flexibility and isolation. The project leverages a modular stack that integrates seamlessly with existing developer tools:

• Webhook deployment: Hosted on Google Cloud Platform (GCP), ensuring scalability and reliability for incoming task triggers.
• Sandbox environments: Tasks run in isolated sandboxes powered by either GCP or Blaxel, providing secure, ephemeral execution spaces.
• Task sourcing: Pulls coding tasks directly from Linear, ensuring context-rich requests are properly addressed.
• Code hosting & CI/CD: Integrates with GitHub for repository access, test execution, and automated pull request generation.

The open-source nature of Broccoli means teams can customize the harness to fit their specific workflows, whether they prefer cloud-based execution or hybrid setups. The project’s GitHub repository provides full documentation, configuration examples, and contribution guidelines for developers looking to adapt it for their own use cases.

Why open source a coding harness?

The Broccoli team believes that organizations relying heavily on coding as a core business function should invest in their own execution harnesses rather than relying solely on generic cloud-based coding agents. By open-sourcing Broccoli, they aim to democratize access to a workflow that prioritizes safety, scalability, and efficiency.

In a landscape crowded with AI tools that promise automation but deliver fragmented workflows, Broccoli positions itself as a cohesive alternative. It doesn’t replace developers—it augments their capabilities by handling the repetitive, context-heavy parts of coding tasks, leaving humans to focus on design, review, and strategic decisions.

As AI coding agents evolve, tools like Broccoli highlight a key trend: the future of development lies not just in smarter models, but in smarter workflows that bridge the gap between automation and human oversight. Whether it becomes a standard in developer toolkits remains to be seen, but its approach offers a compelling glimpse into what’s possible when execution is decoupled from local environments.