Background: Why Now? Why CatPaw?

In 2024, the AI IDE wave led by Cursor swept the globe, but it faced three major pain points domestically:

• High Cost: $20/month subscription fee + opaque token consumption;
• Unstable Network: Access to models like Claude is restricted, often requiring manual proxy configuration;
• Weak Chinese Support: Significant prompt understanding deviations requiring repeated adjustments.

Meituan chose to take a latecomer approach:

✅ Internally stable for 2 years (launched as an editor plugin in 2023) ✅ Public beta officially opens in November 2025, launching the standalone product — CatPaw.

It does not aim to “reinvent the wheel” but instead uses Cursor as the interactive base + deep localization enhancements, carving out a pragmatic, efficient, developer-friendly technical route.

▲ Left: CatPaw | Right: Cursor | Highly consistent interaction logic, zero learning cost for old users to migrate.

Core Architecture: Dual-Mode Agent + Toolchain Loop

The core of CatPaw is the Ask / Agent dual-mode, covering all scenarios from lightweight Q&A to fully automated development:

Behind the Agent mode is a complete toolchain loop, currently built-in with 11 types of native tools:

1 File: readFile / writeFile / createDir

2 Code: grep / searchSymbol / findReferences

3 Terminal: runCommand (supports auto-run switch)

4 Network: fetchURL / searchWeb

5 Testing: runTests / debug

More critically, it supports MCP (Model Context Protocol) extensions, for example, the Chrome DevTools MCP can achieve seamless connections from “development → launch → automated testing”.

Knowledge Enhancement: @Docs Indexing System in Action

CatPaw replicates and optimizes Cursor’s signature @Docs feature — teaching AI about your private documents.

Operation Process:

1. Go to Settings → Indexing & Docs;
2. Add URL or local documents (supports Markdown/PDF);
3. Wait for the small green dot ✅ indexed prompt to take effect;
4. Reference knowledge in conversation using @doc_name.

✅ Tested: Indexing both CatPaw official documentation and Cursor Docs simultaneously, the Agent can accurately compare similarities and differences.

The Agent uses LongCat-Flash (Meituan’s self-developed programming training model) to quickly generate structured analysis, producing high-quality outputs:

Developer Experience: Key Innovations

Custom Agent: Your Personal Programming Partner

In Agent → Add Mode, you can create a personalized Agent:

• Specify model (LongCat-Flash / GLM-4.6 / Qwen2.5-Coder);
• Select required toolset (principle of least privilege);
• Write role-based prompts (e.g., “You are a senior React performance engineer”);
• Enable auto-fix (automatically fix Linter) or auto-run (automatically execute terminal commands).

Conversation Favorites: Preserving Quality Context

❗ Industry First: Mainstream AI IDEs (including Cursor, Trae) lack this feature.

When multiple rounds of conversation yield an ideal result, you can ⭐️ favorite the current session — quickly reuse it later in History → Favorites.

Scene Value: Team knowledge reuse, archiving complex problem solutions, preserving prompt templates.

Built-in Unit Testing & Real-time Preview

• Testing Integration: The Agent can generate/run Jest/Pytest with one click, covering mainstream frameworks;
• Real-time Preview: Web projects auto-render, supporting click element → Edit → Agent modification, significantly boosting front-end debugging efficiency.

Built-in System Prompt

You can configure fixed content from conversations as a System Prompt, which CatPaw will automatically send to the model at the start of each sidebar and inline conversation.

For instance, if you are a front-end developer, you can add your team’s coding standards to the System Prompt, ensuring the AI model adheres to your coding style when generating code.

Model Strategy: Domestic First, Open Compatibility

CatPaw currently supports three types of model access:

Practical Application: Develop a “Pictionary” Game from Scratch in 10 Minutes

Requirement:

Next.js front-end drawing + back-end calling GLM-4.5V multimodal API for image recognition + automated testing

Complete Execution Chain of the Agent:

1. Read the user-provided GLM-4.5V_API.md;
2. create next-app;
3. Install the konva drawing library;
4. Write /api/guess/route.ts to encapsulate GLM API;
5. Build the front end: canvas + submit button + results area;
6. npm run dev;
7. Start the service;
8. Call Chrome DevTools MCP, automatically open the browser and take a screenshot for verification.

✅ Completed in 1 conversation (consuming 1 quota), successfully without rework.