zapat

Zapat

On-demand AI dev teams, powered by Claude Code.

Label a GitHub issue. An AI team assembles itself -- engineers, security reviewers, product managers, UX critics, DevOps engineers, QA engineers, and more -- implements the feature, tests it, reviews its own work, and opens a merge-ready PR. When the work is done, the team disbands. No scheduling. No idle seats.

 GitHub Issue                                                    Merged PR
 (you add a label)                                               (automatic)
     |                                                              ^
     v                                                              |
 +--------+    +-----------+    +--------+    +--------+    +-------+
 | Triage |--->| Implement |--->|  Test  |--->| Review |--->| Merge |
 | team   |    |   team    |    |  team  |    |  team  |    | Gate  |
 +--------+    +-----------+    +--------+    +--------+    +-------+
  assembles     assembles        assembles     assembles     risk-based
  works         works            works          works        auto-merge
  disbands      disbands         disbands       disbands

Each stage spins up a purpose-built team of Claude Code agents. They collaborate in real time -- the builder writes code while the security reviewer audits it, the product manager validates scope, the UX critic checks the experience, and specialized roles (DevOps, QA, program management, technical writing) join as needed. When they converge on a result, the team dissolves and the next stage takes over. Your entire dev pipeline runs itself, from issue to merged PR, in about an hour.

Quick Start

Option A: Install as a Claude Code plugin (recommended)

# 1. Add the marketplace
claude plugin marketplace add zapat-ai/zapat

# 2. Install the plugin
claude plugin install zapat

# 3. Open Claude Code in any project
claude

# 4. Run the setup wizard
/zapat:setup

The setup wizard clones Zapat, configures your repos, agents, and notifications, and starts the pipeline. Once running, label any GitHub issue with agent and Zapat handles it.

Option B: Clone and run locally

# 1. Clone the repo
git clone https://github.com/zapat-ai/zapat.git
cd zapat

# 2. Open Claude Code
claude

# 3. Run the setup wizard
/zapat

The /zapat skill walks you through the same setup process from inside the cloned repo.

Requirements

Tool	Version	macOS	Linux
tmux	any	brew install tmux	sudo apt install tmux
jq	any	brew install jq	sudo apt install jq
gh (GitHub CLI)	2.0+	brew install gh	Install guide
Node.js	18+	brew install node	sudo apt install nodejs npm
git	2.20+	brew install git	sudo apt install git
Claude Code CLI	latest	npm install -g @anthropic-ai/claude-code	same

Runs on any always-on macOS or Linux machine.

Important: Resource & Cost Considerations

Zapat runs multiple Claude Code agents concurrently -- a single pipeline job can spin up 4+ agents working in parallel. Before deploying, understand what this means for your setup:

Compute: Each Claude Code agent is a separate process. When several jobs run simultaneously (e.g., one issue being implemented while another is being reviewed), your machine may have 8-12+ agents active at once. This will saturate CPU and memory on a laptop or shared workstation. We strongly recommend running Zapat on a dedicated machine -- a Mac Mini, a Linux server, or a cloud instance that isn't used for your daily work.

API usage: Every agent call uses Claude tokens. A typical issue-to-merge pipeline (triage + implement + review + test) consumes significant token volume. Zapat uses a 3-tier model strategy to balance quality and cost:

Tier	Environment Variable	Default	Used For
Lead	CLAUDE_MODEL	Opus	Team leads / orchestrators
Sub-agent	CLAUDE_SUBAGENT_MODEL	Opus	Reviewers, analysts (Task sub-agents)
Utility	CLAUDE_UTILITY_MODEL	Haiku	Test runners, standups, planning

• Claude Code subscription (Max/Team/Enterprise): You get a usage quota rather than per-token billing. Monitor your usage patterns in your account dashboard to ensure you stay within your plan's limits. Zapat's concurrent agents can consume quota quickly.
• Anthropic API (pay-per-token): Set spending limits in your Anthropic Console to avoid surprises.

Either way:

• Configure the 3-tier model strategy in .env or per-project project.env to choose cost-appropriate models (e.g., Opus for leads, Sonnet for sub-agents, Haiku for utility tasks)
• Adjust MAX_CONCURRENT_WORK in .env to limit how many jobs run simultaneously