Labro

2 min read

Aside left

Screenshot from Labro: mobile layout showing recent agent runs with project, source, and outcome indicators Screenshot from Labro: full dashboard table view showing runs with project, source, agent, model, cost, turns, and outcome columns Screenshot from Labro: single run detail panel showing agent summary, actions taken, and token usage Screenshot from Labro: charts view showing daily outcomes and breakdowns by model, task source, and perspective

What?

An orchestrator that runs AI coding agents on a schedule - picking or raising GitHub Issues, prompting an agent, and recording the result directly in the GitHub issue. Configurable: define which labels or conditions trigger pickup, the agent prompt, agent permissions (comment, open a PR, merge), and how often to run.

See a sample live metrics dashboard at labro.rossarnold.uk and source/docs at github.com/rssrn/labro.

I use Labro conservatively, with human review in the loop. But it’s flexible: if you prefer more agent autonomy: just add merge or deploy permissions and configure your prompts accordingly.

Supports Claude Code, OpenCode, and Codex as agent backends. Labro is broadly similar to commercial features emerging in 2026 such as Claude’s Routine Runs, Copilot Cloud Agent and Cursor Cloud Agent, but has no dependency on these features.

Why?

Labro’s configurable asynchronous use of coding agents supports:

  • Produce AI-assisted output in a measured, prioritised stream of work to maintain a reviewable cadence. Automatically picking all possible work alongside a human-in-the-loop policy can create a bottleneck on the human side.
  • Get occasional suggestions using a random perspective to shake up your project!
  • Let AI subscriptions or headless-only subscription budget (example: Agent SDK Credit) keep working 24/7
  • Low-friction way to use free or free-tier tokens across multiple providers to monitor and progress your project
  • Configure an appropriate model (=cost) per task complexity - leverage free models for simpler tasks
  • Fall back to alternative models if your first choice is not available or budget exceeded. Or just retry later automatically.

Tech

  • Python 3.12 / Docker / SQLite / GitHub Actions or system cron
  • Agent backends: Claude Code CLI, OpenCode, Codex
  • Metrics dashboard: React 18 + Vite + TypeScript / sql.js (SQLite WASM) / Apache ECharts / Cloudflare R2