CC Switch v3.16.1

Vibe Coding (or ‘thought programming’) is all the rage lately. Is everyone heavily using various AI programming CLI tools like Claude Code, Gemini CLI, or OpenCode? They’re great, but switching API providers or swapping to a backup model when the main one is rate-limited can be a real hassle!

CC Switch is a specializedopen-source unified management and visualization platformbuilt for AI programming tools. Think of it as a ‘provider switcher + configuration management panel + MCP/Skills manager’ for tools like Claude Code, Codex, and Gemini CLI.

Features

Let me outline the most genuinely useful features based on real-world usage frequency:

1. Unified Multi-App Management

One interface directly manages Claude Code, Claude Desktop, Codex, Gemini CLI, OpenCode, OpenClaw, and Hermes Agent. Configure your API in one place, and all software syncs automatically.

2. Provider Management (Most Used Feature)

Built-in 50+ provider presets, including official ones like Anthropic, OpenAI, Google, and services accessible in China like Silicon Flow, DeepSeek, Zhipu GLM, MiniMax, Kimi, and PackyCode. Just paste an API Key to import—no need to manually construct Base URLs.

3. Local Routing + Protocol Translation (The Truly Hardcore Part)

This is what sets CC Switch apart from a basic ‘config switcher’. Since v3.16.0, it can perform protocol conversion between OpenAI’s Responses API ↔ Chat Completions API.

What does this mean? Codex natively only understands the Responses API, but most Chinese service providers (DeepSeek, Kimi, GLM, Silicon Flow) offer the Chat Completions interface. Previously, changing the base_url would always result in 404 or 400 errors. Now, CC Switch runs a local proxy at 127.0.0.1:15721,translating incoming requests into a format the provider understands and translating responses back into a format Codex recognizes, all seamlessly.

⚠️ This means you can use DeepSeek in Codex, or GPT in Claude Code—models and applications are decoupled.

4. Automatic Failover + Circuit Breaking

Automatically switches to a backup if the primary provider fails. Triggers a circuit breaker after N consecutive failures to prevent cascading failures. For someone like me who relies on AI as a primary productivity tool, this is the most critical feature.

5. Unified MCP / Prompts / Skills Management

Configure an MCP Server once, and it syncs to four applications: Claude Code, Codex, Gemini CLI, and OpenClaw. Skills can be installed with one click from GitHub repos and toggled on/off per application.

6. Usage & Cost Tracking

Token consumption, request counts, cache hit rates, cost breakdowns by model and provider—this data is especially important when switching between multiple AI programming CLIs.

7. Cloud Sync

Point the config directory to Dropbox / OneDrive / iCloud / WebDAV to keep your settings when switching computers.

Tutorial

If you’re using it for the first time, follow this 5-minute workflow:

Step 1: Download & Install

Go to the GitHub Releases page and download the appropriate installer:

• Windows: choose .msi (supports auto-updates)
• macOS: recommended brew tap farion1231/ccswitch && brew install --cask cc-switch
• Linux: choose .deb/ .rpm/ AppImage based on your distro

⚠️ Windows SmartScreen might block installation; click ‘More info’ → ‘Run anyway’. On macOS, you’ll need to allow the app in ‘Privacy & Security’ on first launch.

Step 2: Go to Settings, Enable Plugin Takeover

Top-left gear icon → General → Turn on ‘Apply to Claude Code / Codex / Gemini CLI plugins’. If you skip this, switching providers won’t work later.

Step 3: Add a Provider

Top-right + → Select a preset (e.g., ‘DeepSeek’, ‘Zhipu GLM’, ‘Silicon Flow’) → Paste API Key → Save.

Step 4: Enable

Click ‘Enable’ in the provider list. Takes effect immediately in Claude Code; Codex/Gemini might require a terminal restart.

Step 5 (Optional): Enable Routing

Settings → Advanced → Routing Service → Turn on routing for the applications you need. This unlocks failover, usage stats, and hot-swapping.

Verification is simple: in the terminal, run claude or codex, send a test request. If you get a reply, it’s working.