Mergeyard

Label a GitHub issue ready-for-agent and let Mergeyard turn it into a reviewed, CI-green pull request. It runs Claude Code or Codex on your machine, gives each issue an isolated worktree, and coordinates implementation, independent review, and fixes. Follow progress in the local dashboard. You decide what merges.

See it in action

Example data in the real dashboard, with time between stages shortened.

Install

Install on macOS or Linux with:

curl -fsSL https://github.com/rcpassos/mergeyard/releases/latest/download/install.sh | sh

Binaries include the dashboard. See installation for PATH setup, checksums, upgrades, version selection, and building from source.

Quickstart

Have Git, GitHub CLI, tmux, and your chosen agents installed and authenticated. Setup checks dependencies and helps configure Claude Code or Codex independently for the implementer and reviewer.

gh auth login
mergeyard init
# Choose owner/repo and accept creation of the queue labels when prompted.

gh issue edit 123 --repo owner/repo --add-label ready-for-agent
mergeyard start

Replace owner/repo and 123 with your repository and issue number. Init runs checks; resolve any reported errors before starting. Start opens the local dashboard and prints a single-use link. Follow the run to Waiting for your merge, inspect its PR, and merge it yourself. See setup if you need help getting started.

How it works

To try the loop in a disposable private repository, build from the current source and run mergeyard try (not included in v0.1.0). It asks before creating the playground, supplies a sample issue and CI, and uses a separate configuration and workspace. See CLI operations for prerequisites and cleanup.

flowchart LR
    Issue["Label an issue ready-for-agent"] --> Implementer
    Implementer --> Reviewer["Independent reviewer"]
    Reviewer -->|Blocking findings| Fix["Implementer fixes"]
    Fix --> Reviewer
    Reviewer -->|Approved| CI
    CI -->|Failed checks| Fix
    CI -->|Green| You["You merge"]

Fix and review rounds are bounded. When a run needs help, use the dashboard to stop it or take over the implementer's conversation. Work stays in its branch, worktree, and draft PR. Scheduling and recovery details.

Permissions

Agents run with your local permissions. The default Claude configuration uses bypassPermissions for unattended execution. Only queue issues and repositories you trust, and review the permission settings before starting.

Documentation

Start with the architecture overview for contributor orientation and links to the current behavior docs.

• Commands and dashboard
• Configuration and example config
• Usage limits and credits
• Development and website
• Releases