CCG (Claude + Codex + Gemini Multi-Model Collaboration) is a multi-model collaboration workflow engine that runs on top of Claude Code. Its positioning is to make Claude a multi-model orchestrator: Claude retains lead control, dispatching specialized work via a Go binary bridge to external models such as Codex (OpenAI), Grok (xAI), Kimi Code (Moonshot), and Antigravity for parallel analysis and review.
Purpose and How It Works Start with a single command: /ccg:go. The user describes what they want to do, and the engine handles the rest—analyzing intent, selecting a strategy, managing the entire workflow. The Hook engine injects state into each round of the Claude Code session, so that context still exists after compaction, and subagents carry their own specifications in prompts, reducing state loss. codeagent-wrapper (a compiled Go binary) is responsible for bridging Claude to external models for parallel analysis and review.
Built-in Strategies The engine automatically selects the appropriate strategy based on task type and complexity, 10 in total. Simple tasks run quickly with zero overhead, while complex tasks enable the full engine.
Hook Engine Injects state into every Claude Code session via 4 JavaScript hooks, allowing context to survive across compaction, with subagents carrying specifications directly in prompts.
Task System Tasks of medium complexity and above receive a persistent directory for lifecycle management.
Quality Gates and Domain Knowledge Built-in security and quality gates. When topics such as security, caching, RAG, Kubernetes, etc. are mentioned in messages, the relevant domain knowledge files are automatically injected, covering 10 domains and 61 files: security, architecture, DevOps, AI/MLOps, development, frontend design, infrastructure, mobile, data engineering, orchestration.
Standalone Skills Provides self-contained skills that can be used independently or automatically triggered by intent, drawn from real website-building work and accompanied by scripts, such as bt-panel (reads credentials only from environment variables or a git-ignored sites.json, with no hardcoding) and seo-page-builder (adapted from the skill of the same name by yuzeiki, see its SKILL.m for detailsd). Native skills can be installed independently starting from v3.7.0, or selected in the CCG installation wizard; each selected skill gets a ccg- entry, and its scripts and reference materials are stored in a versioned source cache. Claude Code uses /ccg-verify-quality, Codex uses $ccg-verify-quality. The manager only exposes selected entries, keeps companion resources together, rejects ownership conflicts, and preserves edited caches. The menu includes S. Manage Skills and V. Doctor. Reinstalling preserves the latest selection; --skills all selects the entire directory, and --skills none selects no skills. Existing full installations can still be used as compatibility mode.
Commands Core (13 commands by default in v3.0), OpenSpec integration, and legacy mode (18 additional commands). Legacy mode includes /ccg:workflow, /ccg:plan, /ccg:execute, /ccg:frontend, /ccg:backend, /ccg:analyze, /ccg:debug, /ccg:optimize, /ccg:test, /ccg:review, /ccg:team, etc.
Installation Methods Option A — Full installation (multi-model orchestration + commands + hooks + binaries, recommended): requires Node.js 20+ and Claude Code CLI; Codex CLI, Grok CLI, Kimi Code CLI, and Antigravity are optional (used to enable multi-model features). The command is npx ccg-workflow. Option B — Native plugin (skills only, zero dependencies, v3.5.1+): use when you only want skills (web-ops toolkit, quality gates, frontend design, domain knowledge) and do not need multi-model orchestration. Skills are invoked with /ccg:<skill>, require no npx, and require no binaries. Multi-model commands are provided only through Option A (their templates contain install-time placeholders that the native plugin cannot resolve). Environment variables are set under "env" in ~/.claude/settings.json.
OpenCode and DeepSeek Working Mode You can configure providers in OpenCode with /connect, then via opencode models (or opencode models deepseek) to query its actual model ID; in CCG's model menu, select OpenCode and enter that provider/model. Claude Code still acts as the coordinator, and the worker runs through the existing wrapper. CCG preserves the selected model when changing routes, reinstalling, or updating, but does not manage the provider's credentials.
CCG on DeepSeek Harness (dsh-ccg) The same role matrix runs natively inside DeepSeek Harness, shipped with this package, with no second installation or tracking of a second version. You can select D. DeepSeek Harness from npx ccg-workflow. Requires dsh 0.2.0-rc.2 or 0.2.1-alpha.2; after updating CCG, you need to run npx ccg-workflow@latest dsh install again and restart dsh. Configuration now belongs to each profile's cordis.patch.yml; the CCG card is located in Settings › Built-in plugins.
Seven roles with fixed delegation tools Each tool resides on its own model and carries its own expert persona, and also provides two capabilities that are harder to achieve on the Claude Code side: the model panel (giving one role multiple models, each answering the same brief independently and presented side by side in the conversation, with no voting and no averaging—the disagreement itself is the finding); and real-time teammates (ccg_team hires a role as a colleague that stays alive across turns, has its own files, rejects conflicting hires rather than warning, and proactively reports back; every hire requires your approval first). No external CLI, no binary bridge, no cold-start overhead—every hop is a provider API request. The source code is located in the dsh-ccg/ directory.
doctor self-check doctor --json returns a versioned report containing check IDs, status, remediation suggestions, and failure/warning counts. It can detect the actual Claude/Codex/skills-only mode, and validate the selected backend (including review routing), Hook registration, and managed skills. It exits with exit code 1 on failure. Version and filesystem checks are used to establish configuration health; they do not call models and do not prove vendor API accessibility. Standalone Codex mode checks its own CLI, configuration, and installed files; a missing shared Claude route is optional, and missing external delegation tools are reported as warnings with setup instructions. Model operation Hooks now follow the configured route and specific model ID, avoid duplicate calls to the same backend, and use native Agents in a pure Claude environment. Native mode uses on-demand skill discovery; legacy knowledge excerpts have a fixed size limit.
Other notes The author also maintains the DSH Marketplace (a DeepSeek Harness plugin directory with 14 categories, 2500+ entries, bilingual), as well as a Python SDK and public API. Project documentation is available in Simplified Chinese, English, and a Documentation version. Installation requires Node.js 20+ and the Claude Code CLI.