Multi-Agents Plugin for Claude Code
A Claude Code plugin that orchestrates configurable AI agents in parallel for code reviews and panel discussions. Supports Codex, Gemini, OpenCode, Pi, Qwen, and any future CLIs via the agent catalog.
Skills
| Skill | Command | Description |
|---|
| ask | /ask <question> | Ask all configured agents a question in parallel and display their responses with a brief synthesis |
| fix-pr | /fix-pr <PR> | Iterative multi-agent PR fix — reviews a PR with multiple agents, fixes critical/major issues, commits, and re-reviews until clean |
| review-pr | /review-pr <PR> | Multi-agent PR review — launches configured agents in parallel to review a GitHub PR, then synthesizes findings into a single comment |
| round-table | /round-table <topic> | Multi-agent panel discussion — configured AI panelists with distinct roles debate a topic across multiple rounds |
Configuration
Configure which agents to use via a ## Multi-Agents section in your CLAUDE.md (project-level or user-level):
## Multi-Agents
- codex
- gemini
- opencode: bailian-coding-plan/glm-5
- opencode: bailian-coding-plan/kimi-k2.5
- pi: bailian-coding-plan/qwen3.5-plus
- qwen
Syntax: - {cli-name} or - {cli-name}: {model}
- Same CLI can appear multiple times with different models
- Lines not matching the pattern are ignored
- Project CLAUDE.md takes precedence over user CLAUDE.md
Skipping Agents
Use --skip to exclude specific agents from a single invocation:
/review-pr 42 --skip codex # skip Codex for this review
/ask --skip glm-5 "your question" # skip the GLM-5 agent
/round-table --skip codex --skip gemini "topic" # skip multiple agents
Matches by CLI name, model name, or display name (case-insensitive).
Auto-Detection Fallback
If no ## Multi-Agents section is found, the plugin auto-detects available CLIs:
which codex gemini opencode pi qwen
One default entry (no model override) is added for each CLI found on $PATH. Skills degrade gracefully — if a tool is missing, the skill skips it and continues with the others.
Supported Agents
| Agent | Binary | Install | Notes |
|---|
| Codex | codex | npm install -g @openai/codex | OpenAI's coding agent |
| Gemini CLI | gemini | npm install -g @anthropic-ai/gemini-cli | Google's Gemini agent |
| OpenCode | opencode | See repo for install | Multi-model agent (supports -m model flag) |
| Pi | pi | brew install pi or see repo | Multi-model agent (supports --model provider/model flag) |
| Qwen Code | qwen | npm install -g @anthropic-ai/qwen-code | Alibaba's Qwen agent (supports -m model flag) |
See references/agent-catalog.md for shared conventions and per-agent files (references/{cli-name}.md) for command templates, quirks, and full CLI details.
Prerequisites
| Tool | Install | Purpose |
|---|
| GitHub CLI | brew install gh | Required by review-pr for PR operations |
Plus at least one configured or auto-detected agent from the supported list above.
Installation
claude plugin marketplace add quabug/multi-agents-plugin
claude plugin install multi-agents@multi-agents-plugin
Usage
Fix PR
Review a GitHub PR with multiple AI agents, then iteratively fix the issues found:
/fix-pr 42
/fix-pr https://github.com/owner/repo/pull/42
What it does:
- Resolves the agent roster (from config or auto-detection)
- Checks out the PR branch and fetches existing review comments
- Launches all configured agents in parallel to review the current diff
- Synthesizes findings by severity — presents critical/major issues to user for confirmation
- Fixes all critical and major issues (including unresolved human reviewer feedback), then commits
- Re-reviews the updated diff — repeats until clean or max 3 iterations
- Reports a summary of iterations, issues fixed, and commits made
Key differences from review-pr:
- Fixes code instead of posting a review comment
- Iterative — reviews, fixes, re-reviews in a loop
- Includes human feedback — existing PR comments are treated as first-class issues to resolve
- Auto-fix — fixes critical/major issues automatically without prompting
- Pushes after each fix commit — keeps the remote PR up to date so reviewers can see incremental fixes
PR Review
Review a GitHub PR with multiple AI agents in parallel:
/review-pr 42
/review-pr https://github.com/owner/repo/pull/42
quabug