CCG Workflow: Claude as the Orchestrator for Codex, Gemini, and Other Model Teams
多模型协作工作流引擎 — /ccg:go 一个命令，AI 自动分析意图、选择策略、编排 Codex + Gemini + Claude 协作执行
At a glance
- What is it?
- CCG is a Go-bridged workflow engine that keeps Claude Code in charge while dispatching tasks to Codex, Grok, Kimi, and Antigravity. It targets developers who want multi-model collaboration without leaving their Claude terminal.
- Who is it for?
- Adopt CCG if you live inside Claude Code and want parallel model analysis or review without switching terminals. It suits teams already paying for multiple model APIs and needing structured task files.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 1 day ago.
- What is it written in?
- Mainly Go, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The Problem: One Model Is Not Enough, but Managing Many Is a Mess
The repository is written in Go, with a Node.js wrapper for installation. It is open source under the MIT license, so you can modify the bridge or the engine. The project also ships a related plugin directory called DSH Marketplace, which is a separate effort by the same author. That context matters because it shows the author cares about tool discoverability, but it is not part of CCG itself.
Architecture: Claude Leads, a Go Binary Bridges the Rest
This is a deliberate design. The hard stop before implementation is not a courtesy; it is a safety valve. The README shows that the engine can parallelize analysis, but it does not let the models run wild. That separation between planning and execution is what makes CCG different from a simple model router.
Installation and First Run: One Command, Then a Menu
If you are not using DeepSeek Harness, you can ignore that part. But the existence of a dsh subcommand suggests the project is actively maintained for that environment. The Go binary codeagent-wrapper is distributed as pre-built binaries, per the release named 'preset'. That means you do not compile Go yourself; you download a binary. That is convenient but raises a question about platform coverage. The README does not list supported operating systems, so verify that a binary exists for your OS before relying on it.
DeepSeek Harness Integration: Model Panels and Live Teammates
The README claims these features are things the Claude Code side cannot do as cleanly. That is a direct comparison, but it does not explain why. It says there is no external CLI, no binary bridge, no cold-start tax; every hop is a provider API request. In DeepSeek Harness, the integration is native, so it avoids the overhead of spawning a separate process. For a user evaluating CCG, this section matters because it shows the project is not just a Claude Code plugin; it is a broader orchestration concept that adapts to different hosts.
The Workflow in Practice: From Intent to Task File to Hard Stop
One notable detail is the classification step. The engine does not just pass your request to a model; it assesses complexity and risk before choosing a strategy. That is a form of meta-reasoning that adds latency but can prevent overspending on trivial tasks. The task file under .ccg/ is a concrete artifact you could inspect or version-control. That is a practical advantage: you can see what the engine decided and why, rather than trusting a black box.
Limitations and Wrong Use Cases: Not a Replacement for a Single Model
A wrong use case is a simple task that a single model can handle in one shot. If you just want to rename a variable or fix a typo, the overhead of reading git status, classifying risk, and launching parallel analysis is overkill. The engine's own classification might route simple tasks to a lighter strategy, but the README does not specify what those strategies are. Another wrong case is a team that lacks API keys for the external models. CCG depends on Codex, Gemini, and others being reachable. If you only have a Claude subscription, the orchestration has nothing to dispatch to.
Alternatives: Direct Multi-Agent Frameworks vs. CCG's Claude-Centric Approach
The README also mentions DeepSeek Harness as a host. If you are already in that environment, CCG's native integration may be more efficient than a generic framework that requires a separate bridge. But for a developer outside Claude Code and DeepSeek Harness, a framework like CrewAI might be more portable.
Maintenance and Upgrade Cost: What the Material Shows
The license is MIT, which allows commercial use and modification without copyleft obligations. That is a permissive license, so you can embed the workflow engine in your own tools as long as you preserve the copyright notice. The README does not list any contributors or a governance model, so maintenance appears to be primarily the author's effort. For a production dependency, you should check the issue tracker and commit history on GitHub to gauge responsiveness. The README includes a badge for DeepWiki and a sponsor section, which indicates community interest, but those are not quality metrics.
Editorial conclusion
Adopt CCG if you live inside Claude Code and want parallel model analysis or review without switching terminals. It suits teams already paying for multiple model APIs and needing structured task files. Skip it if you prefer a single-model workflow or if your organization blocks external model calls. Before adopting, verify that the Go binary bridge works on your OS and that your API keys for Codex, Gemini, and others are valid. Check the .ccg/tasks directory format against your existing project structure to avoid conflicts.
Community notes