Broccoli: Open-source harness for running AI coding agents from Linear tickets in cloud sandboxes
What Broccoli does
Broccoli turns Linear tickets into shipped PRs by running AI coding agents in isolated cloud sandboxes. Each task gets its own sandbox where it checks out the repo, uses ticket context, implements solutions, runs tests and review loops, then opens a pull request for team review.
Architecture and deployment
The implementation uses:
- Webhook deployment: GCP
- Sandbox: GCP or Blaxel
- Project management: Linear
- Code hosting & CI/CD: GitHub
It's deployed to your own GCP project against your Postgres database with no third-party control plane. The deployment is production-grade using Cloud Run + Secret Manager + webhook deduplication + durable job state.
Key features
- Linear ticket → reviewable PR workflow
- Your infrastructure, keys, and data stay within your tenancy
- AI code review on every PR using Claude and Codex
- Customizable prompt templates that you can fork, tune, and version
- Deploys in ~30 minutes with one bootstrap script and one config file
Deployment prerequisites
To deploy, you need:
- GCP project with billing attached and
gcloudlogged in - GitHub App with specific permissions (Contents, Pull requests, Issues = read/write; Metadata = read-only; subscribed to Pull request review event)
- Linear bot user (dedicated, not personal account) with API key
- OpenAI and Anthropic API keys with billing enabled
- Linear webhook pointing at
${Service URL}/webhooks/linearusing auto-generatedbroccoli-oss-linear-webhook-secret
Usage results
In the last four weeks, 100% of PRs from non-developers were shipped via Broccoli. For developers on the team, this share is around 60%, with more complicated features requiring manual back-and-forth design with Codex/Claude Code.
📖 Read the full source: HN AI Agents
👀 See Also
Agent Kernel: Three Markdown Files for Stateful AI Agents
Agent Kernel provides three markdown files that enable stateful behavior in AI coding agents without databases or custom frameworks. It works with OpenCode, Claude Code, Codex, Cursor, Windsurf, and similar tools.
Automated Claude Code Pipeline Cuts Token Usage from 78k to 15k Per Feature
An open-source pipeline for Claude Code automates 12 phases including pre-check analysis of existing code, reducing token usage from ~78k to ~15k per feature. It offers three profiles (yolo, standard, paranoid) and replaces confidence scores with grep-based validation.
Scrapling integrated as OpenClaw's scraping backbone
Scrapling, an open-source library that learns page structure and adapts to changes, has been integrated into OpenClaw as its core scraping engine. It's 774x faster than BeautifulSoup with Lxml and supports multiple selector types with async sessions.
A Pattern for Running Claude Code on Overnight Unattended Sessions Without Drift
A three-piece framework — chain runner, supervisor, and a single handoff contract — solves the feedback-loop drift problem in multi-hour autonomous Claude Code sessions.