English | 中文
Website | Online Optimizer | Prompt Garden | Docs | Quick Start | Chrome Extension | 💖 Support
First-time Tutorial | Development Docs | Vercel Deployment Guide | Cloudflare Deployment Guide | MCP Deployment Guide | DeepWiki Docs | ZRead Docs
Prompt Optimizer is a powerful AI prompt optimization tool that helps you write better AI prompts and improve the quality of AI outputs. It supports four usage methods: web application, desktop application, Chrome extension, and Docker deployment.
Prompts can start from manual writing, templates, local imports, or sources such as Prompt Garden. Prompt Optimizer is where those prompts are optimized, tested, evaluated, and saved as reusable prompt assets.
1. Prompt Optimization: Clearer, Better-Structured Output
Clarify goals, constraints, and output structure. Compare both prompt versions on the same model to see changes in organization and emphasis.
2. Variable Templates: Optimize Once, Reuse Across Inputs
Turn changing inputs into variables while reusing instructions and constraints. Compare prompt versions side by side for concise replies that follow your requirements.
3. Arena: One Task, Multiple Models
Give the same prompt to multiple models. Compare output quality, style, and instruction following side by side to choose a model for your task.
- 🎯 Intelligent Optimization: One-click prompt optimization with multi-round iterative improvements to enhance AI response accuracy
- 📝 Multiple Optimization Modes: System prompts, user prompts, variable templates, and multi-message context
- ⚖️ Arena Comparisons: Compare models or parameter settings side by side on the same task
- 🔄 Analysis and Compare Evaluation: Supports analysis, single-result evaluation, and multi-result compare evaluation to help determine whether a prompt has truly improved
- 🤖 Multi-model Integration: Support for mainstream AI models including OpenAI, Gemini, DeepSeek, Grok, Zhipu AI, SiliconFlow, MiniMax, etc.
- 🖼️ Image Generation: Support for Text-to-Image (T2I), Image-to-Image (I2I), and Multi-Image generation with models like Gemini, Seedream, Grok
- 🌱 Prompt Sources: Start from manual writing, templates, local imports, or Prompt Garden import codes
- ⭐ Smart Favorites: Resource-aware prompt assets with version history, reproducible examples, media support, source binding, and workspace application
- 📊 Advanced Testing Mode: Context variable management, multi-turn conversation testing, Function Calling support
- 🔒 Secure Architecture: App data is stored locally by default; model requests go to your configured service
- 📱 Multi-platform Support: Available as web application, desktop application, Chrome extension, and Docker deployment
- 🔐 Access Control: Password protection feature for secure deployment
- 🧩 MCP Protocol Support: Supports Model Context Protocol (MCP), enabling integration with MCP-compatible AI applications like Claude Desktop
- 🖼️ Text-to-Image (T2I): Generate images from text prompts
- 🎨 Image-to-Image (I2I): Transform and optimize images based on local files
- 🖼️ Multi-Image Generation: Use multiple input images to constrain subject relationships, sequential semantics, and final generation goals
- 🔌 Multi-model Support: Integrated with mainstream image generation models like Gemini, Seedream, Grok
- ⚙️ Model Parameters: Support model-specific parameter configuration (size, style, etc.)
- 📥 Preview & Download: Real-time preview of generated results with download support
- 🔄 Style Transfer: Learn style, composition, and color from reference images
- 🌱 Optional Prompt Sources: Bring prompts from manual writing, templates, local files, or Prompt Garden
- 📥 Import & Collect: Import prompts with metadata, media, examples, and source binding when available
- ⭐ Resource-aware Assets: Save stable prompts as reusable favorites with version history
- 🔗 Source Binding: Track prompt origins and maintain reproducible examples without requiring a specific source
- 📦 Complete Backup: Export and import favorites with all referenced resources
- 📊 Context Variable Management: Custom variables, batch replacement, variable preview
- 💬 Multi-turn Conversation Testing: Simulate real conversation scenarios to test prompt performance in multi-turn interactions
- 🛠️ Function Calling Support: Function Calling integration with support for OpenAI and Gemini tool calling
- 🔍 Analysis and Evaluation Pipeline: Supports analysis, evaluation, compare evaluation, and evaluation-driven smart rewrite in text modes
Image workspace guides: Text-to-Image · Image-to-Image · Multi-Image
Direct access: https://prompt.always200.com
The online app stores its app data in your browser. Model requests go directly to the provider you configure. For your first run, connect one working text model in Model Manager, then follow the first-time tutorial to improve and copy a prompt.
Method 1: One-click deployment to your own Vercel:
Method 2: Fork the project and import to Vercel (Recommended):
- First fork the project to your GitHub account
- Then import the project to Vercel
- This allows tracking of source project updates for easy syncing of new features and fixes
- Configure environment variables:
ACCESS_PASSWORD: Set access password to enable access restrictionVITE_OPENAI_API_KEYetc.: Optional private-deployment model settings. Do not preconfigure API keys on public frontend deployments becauseVITE_*values are exposed in browser assets.
For more detailed deployment steps and important notes, please check:
Use the Deploy to Cloudflare button for the quickest public-repository setup. It creates a repository under your GitHub/GitLab account and deploys with Workers Builds. For private repositories or stricter repository-access control, import your own repository manually; keep the default deploy commands and clear the build command if Cloudflare auto-fills pnpm run build, because wrangler.jsonc builds the Web frontend and publishes packages/web/dist as static assets.
For access control and analytics on Cloudflare, configure Cloudflare Access and Cloudflare Web Analytics in the Cloudflare dashboard. No frontend dependency or application-code change is required.
Download the desktop build for your platform from GitHub Releases. Use the .exe installer for Windows, the .dmg disk image for macOS, or the .AppImage for Linux. Where available, .zip archives can be extracted and run directly.
The desktop app can connect to local models such as Ollama and to API providers that restrict browser requests. The Windows installer and Linux AppImage support in-app updates. On macOS, the app can check for updates. Download and install the new build from GitHub Releases.
- Install from Chrome Web Store (may not be the latest version due to approval delays): Chrome Web Store
- Click the icon to open the Prompt Optimizer
Click to view Docker deployment commands
Set ACCESS_PASSWORD to protect the web interface. ACCESS_USERNAME is optional and defaults to admin.
# Run container (default configuration)
docker run -d -p 8081:80 --restart unless-stopped --name prompt-optimizer linshen/prompt-optimizer
# Run container (with API key configuration and password protection)
docker run -d -p 8081:80 \
-e VITE_OPENAI_API_KEY=your_key \
-e ACCESS_USERNAME=your_username \
-e ACCESS_PASSWORD=your_password \
--restart unless-stopped \
--name prompt-optimizer \
linshen/prompt-optimizerClick to view Docker Compose deployment steps
# 1. Clone the repository
git clone https://github.com/linshenkx/prompt-optimizer.git
cd prompt-optimizer
# 2. Copy the template and edit .env to set your own ACCESS_PASSWORD
# and any model API keys used by the Compose service.
cp env.local.example .env
# 3. Start the service
docker compose --env-file .env -f docker/docker-compose.yml up -d
# 4. View logs
docker compose --env-file .env -f docker/docker-compose.yml logs -fThe included Compose file serves the web app at http://localhost:28081 and MCP at http://localhost:28081/mcp. Edit the Compose file to change the host port or which environment variables are passed to the container.
Click to view MCP Server usage instructions
Prompt Optimizer now supports the Model Context Protocol (MCP), enabling integration with AI applications that support MCP such as Claude Desktop.
When running via Docker, the MCP Server automatically starts and can be accessed via http://ip:port/mcp.
MCP Server requires API key configuration to function properly. Main MCP-specific configurations:
# MCP Server Configuration
MCP_DEFAULT_MODEL_PROVIDER=openai # Options: openai, gemini, anthropic, deepseek, grok, siliconflow, zhipu, dashscope, openrouter, modelscope, custom
MCP_LOG_LEVEL=info # Log levelIn a Docker environment, the MCP Server runs alongside the web application. You can access the MCP service through the same port as the web application at the /mcp path.
For example, if you map the container's port 80 to port 8081 on the host:
docker run -d -p 8081:80 \
-e VITE_OPENAI_API_KEY=your-openai-key \
-e MCP_DEFAULT_MODEL_PROVIDER=openai \
--name prompt-optimizer \
linshen/prompt-optimizerThe MCP Server will then be accessible at http://localhost:8081/mcp.
Use that URL with a client that supports Streamable HTTP. For a local connection in Claude Desktop, see the stdio setup.
- optimize-user-prompt: Optimize user prompts to improve LLM performance
- optimize-system-prompt: Optimize system prompts to improve LLM performance
- iterate-prompt: Iteratively improve mature prompts based on specific requirements
For more detailed information, please refer to the MCP Server User Guide.
Click to view API key configuration methods
- Open Model Manager from the top bar.
- In Text Models, click Add. Select a provider, enter its API key, and select a model your account can use.
- Click Test Connection, then Create.
Set connection defaults under Advanced Parameters in Model Manager. To try a different setting for one test, open that test's model dropdown and adjust the parameters it exposes. See Model Selection and Parameters.
Configure environment variables through the -e parameter when deploying with Docker:
-e VITE_OPENAI_API_KEY=your_key
-e VITE_GEMINI_API_KEY=your_key
-e VITE_DEEPSEEK_API_KEY=your_key
-e VITE_GROK_API_KEY=your_key
-e VITE_ZHIPU_API_KEY=your_key
-e VITE_SILICONFLOW_API_KEY=your_key
# Multiple Custom Models Configuration (Unlimited Quantity)
-e VITE_CUSTOM_API_KEY_ollama=dummy_key
-e VITE_CUSTOM_API_BASE_URL_ollama=http://localhost:11434/v1
-e VITE_CUSTOM_API_MODEL_ollama=qwen2.5:7b📖 Detailed Configuration Guide: See Multiple Custom Models Documentation for complete configuration methods and advanced usage
Use Node.js 24 (24.22.1 or later in the 24.x series) and pnpm 10.6.1. See the development guide for details.
Click to view local development commands
# 1. Clone the project
git clone https://github.com/linshenkx/prompt-optimizer.git
cd prompt-optimizer
# 2. Install dependencies
pnpm install
# 3. Start development server
pnpm dev # Main development command: build core/ui and run web app
pnpm dev:fresh # Complete reset and restart development environment- Basic feature development
- Web application release
- Chrome extension release
- Internationalization support
- Support for system prompt optimization and user prompt optimization
- Desktop application release
- MCP service release
- Advanced mode: Variable management, context testing, function calling
- Image generation: Text-to-Image (T2I) and Image-to-Image (I2I) support
- Prompt favorites and template management
- Support for workspace/project management
For detailed project status, see Project Status Document
- Documentation Index - Index of all documentation
- Technical Development Guide - Technology stack and development specifications
- LLM Parameters Configuration Guide - Detailed guide for advanced LLM parameter configuration
- Project Structure - Detailed project structure description
- Project Status - Current progress and plans
- Product Requirements - Product requirements document
- Vercel Deployment Guide - Detailed instructions for Vercel deployment
- Cloudflare Deployment Guide - Web frontend deployment on Cloudflare Workers / Pages
Click to view frequently asked questions
A: Most connection failures are caused by Cross-Origin Resource Sharing (CORS) issues. As this project is a pure frontend application, browsers block direct access to API services from different origins for security reasons. Model services will reject direct requests from browsers if CORS policies are not correctly configured.
A: Ollama fully supports the OpenAI standard interface, just configure the correct CORS policy:
- Set environment variable
OLLAMA_ORIGINS=*to allow requests from any origin - If issues persist, set
OLLAMA_HOST=0.0.0.0:11434to listen on any IP address
Q3: How to solve CORS issues with commercial APIs (such as Nvidia's DS API, ByteDance's Volcano API)?
A: These platforms typically have strict CORS restrictions. Recommended solutions:
-
Use Desktop Application (Most Recommended)
- Desktop app has no CORS restrictions as a native application
- Can directly connect to any API service, including locally deployed models
- Provides the most complete and stable feature experience
- Download from GitHub Releases
-
Use Self-deployed API Proxy Service (Professional solution)
- Deploy open-source API aggregation/proxy tools like OneAPI, NewAPI
- Configure as custom API endpoint in settings
- Request flow: Browser → Proxy service → Model service provider
- Full control over security policies and access permissions
Note: All web versions (including online version, Vercel deployment, Docker deployment) are pure frontend applications and subject to browser CORS restrictions. Only the desktop version or using an API proxy service can solve CORS issues.
Q4: I have correctly configured CORS policies for my local model (like Ollama), why can't I still connect using the online version?
A: This is caused by the browser's Mixed Content security policy. For security reasons, browsers block secure HTTPS pages (like the online version) from sending requests to insecure HTTP addresses (like your local Ollama service).
Solutions: To bypass this limitation, you need to have the application and API under the same protocol (e.g., both HTTP). We recommend the following approaches:
- Use the desktop version: Desktop applications have no browser restrictions and are the most stable and reliable way to connect to local models
- Use Docker deployment (HTTP): Access via
http://localhost:8081, both the app and local Ollama use HTTP - Use Chrome extension: Extensions can bypass some security restrictions in certain situations
A: This is because the application has not been signed with an Apple Developer certificate. Due to the high cost of Apple Developer accounts, the desktop application is currently unsigned.
Solution: Run the following command in Terminal to remove the quarantine attribute:
# For installed applications
xattr -rd com.apple.quarantine /Applications/PromptOptimizer.app
# For downloaded .dmg files (run before installation)
xattr -rd com.apple.quarantine ~/Downloads/PromptOptimizer-*.dmgAfter running the command, you can open the application normally.
Click to view contribution guidelines
- Fork the repository
- Create a feature branch (
git checkout -b feature/AmazingFeature) - Commit your changes (
git commit -m 'Add some feature') - Push to the branch (
git push origin feature/AmazingFeature) - Open a Pull Request
Tip: When developing with Cursor tool, it is recommended to do the following before committing:
- Use the "CodeReview" rule for review
- Check according to the review report format:
- Overall consistency of changes
- Code quality and implementation method
- Test coverage
- Documentation completeness
- Optimize based on review results before submitting
Thanks to all the developers who have contributed to this project!
This project was partly inspired by LangGPT in prompt engineering and structured prompt design. Thanks to the LangGPT project and community for their open-source sharing and continued exploration.
This project is licensed under AGPL-3.0.
In simple terms: You can freely use, modify, and commercialize this project, but if you turn it into a website or service for others, you must share your source code.
👉 Click for detailed explanation
What you can do:
- ✅ Personal use, learning, and research
- ✅ Internal company use (not offering public services)
- ✅ Modify code for commercial projects
- ✅ Charge for products or services
What you must do:
- 📖 If distributing software or offering network services, disclose source code
- 📝 Preserve original author's copyright notices
Core principle: Commercial use is allowed, but not closed-source.
If this project is helpful to you, please consider giving it a Star ⭐️
- Submit an Issue
- Create a Pull Request
- Join the discussion group
