Agent Integration

RAPID integrates with AI coding CLI tools by wrapping them rather than reimplementing them. This document explains how RAPID manages and orchestrates these tools.

Supported Agents

Agent	CLI Command	Instruction File	Provider
Claude Code	claude	CLAUDE.md	Anthropic
OpenCode	opencode	AGENTS.md	Multi-provider
Aider	aider	.aider.conf.yml	OpenAI/Anthropic
GitHub Copilot	gh copilot	—	GitHub

Integration Model

RAPID CLI takes three inputs:

1. Config (rapid.json) — Which agents, how to install them
2. Secrets (1Password/Vault) — API keys loaded securely
3. Context (AGENTS.md) — Project-specific instructions

These flow through:

1. Environment Setup

   • Export API keys as env vars
   • Generate/update instruction files
   • Configure MCP servers

2. Agent Launcher

   • Install CLI if missing
   • Launch with configured args
   • Manage sessions for multi-agent mode

3. AI CLI Tool

   • Reads environment variables and instruction files
   • Provides chat interface, code generation, file editing

How RAPID Wraps Agents

1. Environment Preparation

Before launching any agent, RAPID:

# Loads secrets and exports as environment variables
export ANTHROPIC_API_KEY="sk-ant-..."
export OPENAI_API_KEY="sk-..."

# Sets working directory
cd /workspaces/project

# Ensures instruction files exist
# (AGENTS.md, CLAUDE.md, etc.)

2. Agent Installation

If the CLI tool is missing, RAPID installs it:

# From rapid.json agents.available.<name>.installCmd
npm install -g @anthropic-ai/claude-code  # Claude
npm install -g opencode                     # OpenCode
pip install aider-chat                      # Aider

3. Agent Launch

RAPID executes the configured CLI:

# Single agent mode
claude

# With additional args (from rapid.json)
aider --model gpt-4o --auto-commits

Instruction Files

Each agent reads project-specific instructions from designated files.

CLAUDE.md

Used by Claude Code to understand project context:

# Project: my-api

## Overview

Express.js REST API with PostgreSQL database.

## Code Style

- TypeScript with strict mode
- Async/await for all async operations
- Zod for validation

## Commands

- `npm run dev` - Start development server
- `npm test` - Run tests
- `npm run build` - Production build

AGENTS.md

Generic instruction file used by OpenCode and others:

# Agent Instructions

## Project Type

TypeScript Node.js application

## Important Files

- src/index.ts - Entry point
- src/config.ts - Configuration
- package.json - Dependencies

## Testing

Always run `npm test` after changes.

.aider.conf.yml

Aider-specific configuration:

model: gpt-4o
auto-commits: true
gitignore: true
map-tokens: 1024

Multi-Agent Mode

RAPID supports running multiple agents concurrently:

rapid dev --multi

This creates a split-pane session with each configured agent:

Pane	Agent	Use Case
Left	Claude Code	Architecture, complex reasoning
Right	OpenCode	Implementation
Bottom	Aider	Quick fixes, auto-commits

Controls

Key	Action
Ctrl+b then o	Switch panes
Ctrl+b then arrow	Navigate panes
Ctrl+b then z	Zoom current pane
Ctrl+b then d	Detach (keeps running)

Adding Custom Agents

Add new agents to rapid.json:

{
  "agents": {
    "available": {
      "my-agent": {
        "cli": "my-ai-tool",
        "instructionFile": "MY_AGENT.md",
        "envVars": ["MY_API_KEY"],
        "installCmd": "npm install -g my-ai-tool",
        "args": ["--verbose", "--model", "custom"]
      }
    }
  }
}

Agent Communication

Agents don’t directly communicate with each other. Instead:

1. All agents share the same filesystem
2. Changes made by one agent are visible to others
3. Use git to track and review changes from any agent
4. Instruction files provide consistent context

Best Practices

1. Use specific agents for specific tasks

   • Claude: Architecture, complex reasoning
   • Aider: Quick code changes, refactoring
   • OpenCode: Multi-model flexibility

2. Keep instruction files updated

   • Update when project structure changes
   • Add new patterns as they emerge

3. Review multi-agent changes carefully

   • Each agent may have different styles
   • Use git diff to review before committing

4. Set appropriate context

   • More context = better results
   • But too much context = slower responses