CodeBridge

Multi-Agent Coding Orchestrator CLI — Coordinate multiple AI coding agents to collaborate on complex tasks

CodeBridge is a lightweight CLI orchestrator that enables multiple coding agents (Claude Code, Codex CLI, Qwen Code, Aider, Gemini CLI) to collaborate on tasks through a structured planner → critic → executor → reviewer loop.

What is CodeBridge?

CodeBridge does NOT replace any coding agent — it sits above them, coordinating their work through their non-interactive/headless modes. The unique value proposition: any coding CLI can talk to any other coding CLI to complete a task together. One agent plans, another executes, a third reviews. The orchestrator manages the conversation between them.

Architecture

┌──────────────────────────────────────────────────────────┐
│                   CodeBridge CLI                         │
│                                                          │
│  ┌────────────────────────────────────────────────────┐  │
│  │              Orchestration Engine                   │  │
│  │                                                    │  │
│  │   PLAN → CRITIQUE → EXECUTE → REVIEW → REPAIR     │  │
│  │                                                    │  │
│  │   State Machine + Run Manager + Config Loader      │  │
│  └────────────────────────────────────────────────────┘  │
│                          │                               │
│  ┌────────────────────────────────────────────────────┐  │
│  │              Adapter Layer (Plugin System)          │  │
│  │                                                    │  │
│  │  ┌──────────┐ ┌──────────┐ ┌──────────┐          │  │
│  │  │ Claude   │ │ Codex    │ │ Qwen     │          │  │
│  │  │ Code     │ │ CLI      │ │ Code     │          │  │
│  │  │ Adapter  │ │ Adapter  │ │ Adapter  │  ...     │  │
│  │  └──────────┘ └──────────┘ └──────────┘          │  │
│  └────────────────────────────────────────────────────┘  │
└──────────────────────────────────────────────────────────┘

Quick Start

Installation

# Install globally from npm
npm install -g @aryanbhargav/codebridge

# Or run without installing
bunx @aryanbhargav/codebridge --help

# Install dependencies
bun install

# Build the project
bun run build

# Link for local development
bun link

# Or run directly
bun bin/codebridge.ts --help

Initialize Configuration

# Create codebridge.config.yaml in your project
codebridge init

Check Available Adapters

# See which coding agents are installed
codebridge adapters

Run Your First Task

# Basic usage
codebridge run --task "Add error handling to src/api.ts" \
  --planner claude-code \
  --executor qwen-code

# With critic and reviewer
codebridge run --task "Refactor authentication module" \
  --planner claude-code \
  --executor qwen-code \
  --critic claude-code \
  --reviewer claude-code

# Auto-approve (skip human confirmation)
codebridge run --task "Fix bug in checkout flow" \
  --planner claude-code \
  --executor qwen-code \
  --auto-approve

Supported Agents

| Agent | Binary | Status | Notes | |-------|--------|--------|-------| | Claude Code | claude | ✅ Full Support | Best for planning and reviewing | | Qwen Code | qwen | ✅ Full Support | Configure model in ~/.qwen/settings.json | | Codex CLI | codex | ✅ Full Support | OpenAI Codex integration | | OpenCode | opencode | ✅ Full Support | Install with npm i -g opencode-ai@latest | | Aider | aider | ⚠️ Partial Support | Install: pip install aider-chat | | Gemini CLI | gemini | ✅ Full Support | Google Gemini integration |

Commands

codebridge run

Run a complete multi-agent coding task.

codebridge run [options]

Options:
  -t, --task <task>              Task description or @file.md
  -p, --planner <adapter>        Planner adapter name (default: claude-code)
  -e, --executor <adapter>       Executor adapter name (default: qwen-code)
  --critic <adapter>             Critic adapter name (optional)
  --reviewer <adapter>           Reviewer adapter name (optional)
  --auto-approve                 Skip human approval step
  --max-plan-rounds <n>          Max plan-critique iterations (default: 3)
  --max-repair-rounds <n>        Max review-repair iterations (default: 2)
  -c, --context <files...>       Additional context files (glob patterns)
  --config <path>                Path to config file
  -v, --verbose                  Show debug output

Other Commands

• codebridge plan — Create a plan without executing
• codebridge resume --run <id> — Resume a paused/failed run
• codebridge status [run-id] — Show run status
• codebridge list — List all runs
• codebridge adapters — List available adapters
• codebridge config — Show current configuration
• codebridge init — Create config file

Configuration

Configuration is loaded from (in priority order, later overrides earlier):

1. Built-in defaults
2. ~/.codebridge/config.yaml (user-level)
3. ./codebridge.config.yaml (project-level)
4. CLI flags

Example Configuration

See codebridge.config.example.yaml for a complete example.

How It Works

CodeBridge follows a structured workflow:

1. PLANNING: Planner agent creates a detailed implementation plan
2. CRITIQUING (optional): Critic agent reviews the plan for gaps and risks
3. AWAITING_APPROVAL: Human reviews and approves the plan (unless --auto-approve)
4. EXECUTING: Executor agent implements the approved plan
5. REVIEWING: Reviewer agent verifies the implementation
6. REPAIRING (if needed): Executor fixes issues identified by reviewer
7. COMPLETED: Task successfully finished

All interactions are logged in .runs/<run-id>/.

Development

# Install dependencies
bun install

# Run in dev mode
bun run dev

# Build
bun run build

# Test
bun test

# Type check
bun run typecheck

Design Principles

1. Thin orchestrator — CodeBridge is <2000 lines of TypeScript
2. Adapters are dumb pipes — Just CLI wrappers, no agent-specific logic
3. Everything is a message — Full auditability and debuggability
4. Human in the loop by default — Auto-approve is opt-in
5. Fail gracefully — Full error context captured
6. No lock-in — Swap agents with a single flag

License

MIT

Built with ❤️ using Bun and TypeScript.