Open Source · Multi-Model · MCP Ecosystem · Built for Everyone
cc-wrap is a desktop wrapper for Claude Code, combining the agent loop engine with a full graphical interface. No more terminal commands — manage file operations, command execution, code search, and tool extensions through a windowed UI.
Deep optimization for global models: Native support for DeepSeek V4, Qwen3.6, Kimi K2.6, Doubao, GLM-5.1, and more. Thinking/reasoning parameters auto-adapted per model — works out of the box.
Every model uses a different "thinking" parameter format. cc-wrap detects the model name and injects the correct format automatically:
| Model | Parameter Format |
|---|---|
| Claude (Opus / Sonnet / Haiku) | thinking.type + budget_tokens |
| OpenAI o-series / GPT-5 | reasoning_effort |
| DeepSeek V4+ | thinking.type |
| DeepSeek V3 | enable_thinking |
| Qwen3 / QwQ | enable_thinking |
| Kimi K2.6 | thinking.type |
| Doubao | thinking.type + budget_tokens |
| Gemini 2.5 | thinkingConfig.thinkingBudget |
| Gemini 3.x | thinkingConfig.thinkingLevel |
| GLM-5.1 / MiMo | Always reasons, auto-skip |
Users just pick "Off / Low / Medium / High" — no need to know the underlying parameter differences.
AI decides which tools to call, collaborating across multiple rounds to complete tasks.
- Send a request, AI autonomously calls Read / Write / Edit / Bash / Glob / Grep… until done
- Streaming output shows real-time thinking and operations
- Model-aware context compression (DeepSeek 500K, Gemini 200K, Claude 120K, default 80K)
- Stuck detection: consecutive failures trigger strategy hints, preventing infinite loops
- Smart tool result truncation to prevent context overflow
- 3-Level Message Queue: now (immediate), next (next round), later (background)
- Code Review 9 Stages: Three-state verification (CONFIRMED/PLAUSIBLE/REFUTED)
- Coordinator Multi-Worker: Parallel execution of multiple subtasks
One client for all major models:
| Model | Protocol |
|---|---|
| Claude (Opus 4.8 / Sonnet 5 / Sonnet 4.6 / Haiku 4.5) | Anthropic Messages API |
| GPT-5.6 / GPT-5.5 / GPT-5.4 / o3 / o4-mini | OpenAI Chat API |
| DeepSeek V4-Pro / V4-Flash / R1 | OpenAI compatible |
| Qwen 3.7 Max / Qwen3.6 / Qwen3-Coder | OpenAI compatible |
| Kimi K2.6 / K3 | OpenAI compatible |
| Doubao-Seed-2.0 | OpenAI compatible |
| MiniMax M3 / M2.7 | OpenAI compatible |
| GLM-5.2 / GLM-5.1 / GLM-Z1 | OpenAI compatible |
| Gemini 3.x / 2.5 | OpenAI compatible / Google API |
| Grok 4.x (xAI) | OpenAI compatible |
| MiMo V2 | OpenAI compatible |
Auto-detects protocol: Anthropic format uses /v1/messages SSE, OpenAI format uses /v1/chat/completions streaming. Non-vision models auto-strip images to avoid 400 errors. Versioned-path endpoints like Gemini (/v1beta/openai) and Zhipu GLM (/paas/v4) are auto-adapted — no manual URL tweaking.
Verify connectivity right after configuring endpoint / key:
- Per-model card "Test" button verifies Base URL + API Key + model ID
- Global config can test the default model; edit dialog shows results inline
- Returns actual latency and the model name from the API response; failures show the concrete error (HTTP status + body) to quickly pinpoint bad keys / endpoint 404s / timeouts
Not every tool call needs to be visible, but important ones must not be missed:
- AskUserQuestion — Auto-expands, shows only options, hides raw code
- Write / Edit — Auto-expands on completion, shows file changes
- Agent sub-tasks — Auto-expands, shows progress
- Read / Grep / Bash etc. — Stays collapsed, clean interface
Supports Model Context Protocol in two transport modes:
- stdio — Local MCP servers (Node.js / Python), subprocess communication
- HTTP/SSE — Remote MCP services (e.g., Tavily Search), paste URL to connect
One-click MCP server installation with 5 transport modes: npm / pip / uvx / http / stdio.
MCP Tool Validation: Auto-validates input parameters (required params, type checking) to prevent call failures.
Ctrl+ to toggle,node-pty+xterm.js`, VS Code-style experience:
- Real PTY process (cmd.exe / PowerShell / Git Bash)
- Drag to resize, process persists across close/reopen
- Run git, install packages, check logs without leaving the window
AI automatically breaks down complex requests into subtasks with real-time progress.
- Status at a glance: completed / in progress / pending
- Click to toggle status, manual intervention supported
AI automatically captures key information, persisting across conversations.
- Auto-extracts tech stack, preferences, project context
- Manual add / manage / delete memories
- Auto-injected into every conversation
Inject domain knowledge into System Prompt for better AI understanding.
- Custom Skills (name + description + prompts + trigger keywords)
- Always-active or keyword-triggered activation
- Vendor onboarding: paste install steps, AI executes and registers the Skill
- 5-second cache: Avoids repeated loading, auto-clears on save
| Dimension | cc-wrap | Terminal CLI | Web UI |
|---|---|---|---|
| Multi-model | ✅ 13+ models, one-click switch | ❌ Claude only | |
| Thinking control | ✅ Auto-adapts per model | ❌ None | |
| Chinese UI | ✅ Native Chinese + English | ❌ English terminal | |
| MCP extensions | ✅ One-click install | ❌ None | |
| Integrated terminal | ✅ VS Code style | ✅ IS the terminal | ❌ None |
| File editor | ✅ Syntax highlight + preview | ❌ External editor | |
| Permission mgmt | ✅ Visual modal | ✅ | |
| Memory persistence | ✅ Auto + manual | ❌ None | |
| Offline use | ✅ Local | ✅ | ❌ |
| Open source | ✅ MIT | ✅ | ❌ |
| Feature | Description |
|---|---|
| Agent loop | Multi-round tool calls, streaming output, model-aware compression, stuck detection |
| 3-Level message queue | now (immediate), next (next round), later (background tasks) |
| Code Review | 9-stage review, three-state verification (CONFIRMED/PLAUSIBLE/REFUTED) |
| Coordinator | Multi-worker parallel execution, task coordination |
| Thinking level | Auto-detect model, inject correct thinking/reasoning parameters, toolbar quick switch |
| File ops | Read / Write / Edit / Glob / Grep, supports text / .docx / .pdf / .xlsx / .csv, auto encoding detection |
| Bash execution | Non-blocking spawn, Git Bash / cmd auto-detect, cancel + timeout, dangerous command filtering |
| Multi-model | 13+ models, Anthropic + OpenAI + Google protocols, vision model auto-detect |
| Model test | One-click connectivity test per model / global config, returns latency + actual model name + concrete error |
| MCP integration | stdio + HTTP/SSE dual mode, one-click install, auto-reconnect, input validation |
| Integrated terminal | Ctrl+` toggle, node-pty real terminal, draggable panel |
| Smart expansion | AskUserQuestion / Write / Edit auto-expand, others collapsed |
| Task panel | Auto task breakdown, progress tracking, state machine validation |
| Memory system | Auto-extract + manual management, cross-conversation persistence |
| Skills | Custom prompt injection, keyword / always-on activation modes, 5-second cache |
| File editor | Syntax highlighting, line numbers, find/replace, Markdown preview, file tree |
| Image recognition | Paste / drag auto-save, vision model + MCP tool dual path |
| Dual theme | Claude warm dark + soft beige light, adjustable font size |
| Bilingual | Instant language switching (Chinese / English, uses common terms like Base URL) |
| Token stats | Per-message ↑↓ display, /cost for full conversation details, contribution heatmap |
| Permission mgmt | Write / Edit / Bash modal confirmation, "always allow" persistence |
| System tray | Minimize to tray, right-click menu for new / settings / show / hide |
| Log viewer | Built-in log panel with search, clear, export, 5MB auto-rotation |
| Conversation mgmt | Create, switch, delete, export as Markdown |
| Slash commands | /help /clear /model /memory /mcp /skill /theme /export /cost etc. |
| Keyboard shortcuts | Ctrl+P open file, Ctrl+S save, Esc stop generation, `Ctrl+`` terminal |
| Failure retry | Red-highlight on API failure + one-click retry |
Download the latest cc-wrap Setup X.Y.Z.exe from Releases and run the installer.
git clone <repo>
cd cc-wrap
npm install
npm run rebuild # Compile native modules (node-pty)
npm start # Dev mode
npm run build # Package NSIS installer to dist/Config file: %APPDATA%/cc-wrap/config.json
- API keys encrypted via Electron
safeStorage - Model list, theme, font, language, work directory, recent projects
- Always-allowed tools, custom system prompt, environment variable injection
MCP config: %APPDATA%/cc-wrap/mcp-servers.json
┌───────────────┐ IPC (contextBridge) ┌──────────────────┐
│ Main Process │ ◄────────────────────────────► │ Renderer Process │
│ (Node.js) │ preload.js whitelist │ (Chromium) │
│ │ │ │
│ ├─ main.js │ │ ├─ app.js │
│ ├─ agent-loop │ terminal-output (push) │ ├─ index.html │
│ ├─ api-client │ terminal-write (invoke) │ ├─ main.css │
│ ├─ tools.js │ terminal-spawn (invoke) │ └─ xterm.js │
│ ├─ tool-exec │ agent-stream-text (push) │ │
│ ├─ mcp-client │ agent-permission (push) │ Terminal Panel │
│ ├─ system-prom│ ... │ └─ xterm.js │
│ ├─ task-queue │ │ │
│ ├─ code-review│ │ │
│ ├─ coordinator│ │ │
│ ├─ logger.js │ │ │
│ └─ node-pty │ │ │
└───────────────┘ └──────────────────┘
Security: nodeIntegration: false, contextIsolation: true, IPC channel whitelisting, CSP default-src 'self'.
Tech Stack: Electron 28 + vanilla JavaScript (no frontend framework)
| File | Path |
|---|---|
config.json |
%APPDATA%/cc-wrap/ |
conversations.json |
%APPDATA%/cc-wrap/ |
memory.json |
%APPDATA%/cc-wrap/ |
skills.json |
%APPDATA%/cc-wrap/ |
skills/<name>/SKILL.md |
%APPDATA%/cc-wrap/skills/ |
mcp-servers.json |
%APPDATA%/cc-wrap/ |
logs/app.log |
%APPDATA%/cc-wrap/ |
pasted-images/ |
%APPDATA%/cc-wrap/ |
- API Key required — Not out-of-the-box; you need your own model API credits
- Local single-machine — No cloud sync or team collaboration
- Model variation — Claude has the strongest agent capabilities; some models have less stable tool calling
npm start # Dev mode
npm run build # Package
npm run rebuild # Recompile native modules
npm test # Run Jest unit tests
npm run lint # ESLint code check
npm run lint:fix # Auto-fix formatting issuesMIT
⭐ Star if you find this helpful. Issues and PRs welcome.