Skip to content
CCG Workflow

How to Use Multiple AI Models Together in Claude Code

A single AI model writing code hits a ceiling fast. Claude excels at reasoning and orchestration but isn’t the best at frontend UI. Codex is strong on backend analysis but weak on styling. Gemini produces beautiful components but struggles with complex business logic.

What if each model did what it’s best at?

CCG Workflow is a multi-model collaboration engine that runs inside Claude Code. One command, three models, automatic strategy selection.

Why Multi-Model Collaboration Matters

Section titled “Why Multi-Model Collaboration Matters”

A typical full-stack feature like “add JWT authentication to this API” involves:

  • Backend: Route middleware, token signing/validation, refresh mechanism
  • Frontend: Login page, token storage, request interceptor
  • Cross-cutting: Code organization, error handling, tests

Asking one model to handle everything produces mediocre results across the board. Multi-model collaboration lets each model focus on its strength.

How CCG Divides the Work

Section titled “How CCG Divides the Work”
ModelRoleStrength
ClaudeOrchestratorIntent analysis, strategy selection, code review, final writes
CodexBackend AnalystAPI design, database logic, business flow analysis
GeminiFrontend AnalystUI components, styling, interaction logic

The critical design decision: external models have no file write access. They return analysis and patch suggestions only. Claude reviews everything before writing to the codebase.

Using It in Practice

Section titled “Using It in Practice”

Install CCG:

Terminal window
npx ccg-workflow

Then in Claude Code:

> /ccg:go add JWT authentication to this API

The engine automatically:

  1. Reads project context (tech stack, file structure, git status)
  2. Classifies: feature / L complexity / backend
  3. Selects strategy: full-collaborate
  4. Launches Codex + Gemini in parallel for analysis
  5. Synthesizes a plan → waits for your approval
  6. Spawns Agent Teams Builders for parallel implementation
  7. Runs quality gates + dual-model cross-review

You typed one sentence. The engine handled everything else.

10 Strategies, Automatically Selected

Section titled “10 Strategies, Automatically Selected”

Not every task needs three models. A simple typo fix uses direct-fix — Claude handles it directly with zero overhead:

ComplexityStrategyModels Used
Simple bugdirect-fixClaude only
Small featurequick-implementClaude only
Medium featureguided-developClaude + 1 external
Complex featurefull-collaborateClaude + Codex + Gemini
Complex bugdebug-investigateClaude + dual diagnosis
Code reviewreview-auditClaude + dual cross-review

The engine matches the strategy to the task. Simple tasks stay fast. Complex tasks get the full engine.

State Persistence with Hooks

Section titled “State Persistence with Hooks”

The biggest challenge in multi-model workflows is losing context. Claude Code’s context window is finite — long conversations get compacted.

CCG solves this with 4 JavaScript hooks:

  • workflow-state.js — Injects current task state every turn
  • session-start.js — Restores full context after compaction
  • subagent-context.js — Injects specs into sub-agents
  • skill-router.js — Auto-loads domain knowledge on keyword detection

Result: task state is never lost, even after hundreds of conversation turns.

Getting Started

Section titled “Getting Started”
Terminal window
npx ccg-workflow

Requires Node.js 20+ and Claude Code CLI. Codex CLI and Gemini CLI are optional — install them to unlock multi-model features.

Detailed guide: Quick Start

Section titled “Related Guides”

CCG Workflow is open source (MIT). Star us on GitHub.