CatPaw: Meituan's full-scenario AI agent platform, from writing code to finishing the job

CatPaw is a full-scenario AI agent platform from Meituan. Its positioning is no longer that of an AI coding assistant: it puts a personal AI workspace and enterprise-grade agent hosting inside one client. It syncs in real time across macOS, Windows, iOS and Android, and can also run entirely in the browser through its Cloud mode. You describe the goal, the AI plans the steps and dispatches several agents in parallel, and hands back something usable such as a spreadsheet, a visual report or code.

Treating CatPaw as an IDE undersells it; treating it as a chat box overstates it. What it actually does is define how far an agent may reach: it can read and write the file system, run terminal commands and drive a browser, all inside an isolated environment, with high-risk actions waiting for your approval. That boundary is what decides whether it can finish a whole task rather than hand you a snippet.

The second thread is domain depth. Meituan has packaged more than a decade of local-services experience into installable experts and skills covering store operations, review management, marketing conversion and fulfilment. Digital employees can also sit inside IM tools such as Feishu and WeCom as virtual colleagues, woken by an @ mention. For non-engineering roles this often matters more than code completion.

Main capabilities

Tab completion and ask-to-code

Press Tab to accept completions while typing, or ask inside the IDE chat and get a snippet or an explanation without switching tools. This is the most familiar layer, and the part that overlaps with conventional coding assistants.

Built-in preview and debugging

Front-end changes are previewed immediately, and an element you dislike can be pushed into the conversation for the AI to adjust, collapsing the edit-refresh-edit loop into a single exchange.

Project-level analysis

A codebase index gives the assistant context across the whole project rather than just the open file, which is what makes its suggestions fit the existing structure and conventions.

Multi-agent collaboration and cloud long tasks

Complex work is decomposed automatically, with several agents running concurrently in separate environments. In Cloud mode tasks keep running 24/7 with the machine off or offline, and a job started on the desktop can be checked from a phone.

Installable experts and skills

An expert marketplace installs AI experts in one click, each bundling several skills and sub-agents. Browser actions can be recorded in one pass and turned into a reusable skill that a team can share.

Two-level sandbox and MCP

Workspace-write and full-access sandbox modes draw different permission boundaries around the agent, and MCP support connects external tools and services so the capability surface can keep expanding.

Cross-session memory and panel layout

The AI remembers preferences, habits and working context across sessions and devices. Sidebar, chat, preview, file and terminal panels can all be resized freely.

Slash commands and desktop pet

Recurring prompts become slash commands invoked with a single character. While a task runs, a desktop pet waits alongside and raises a sign when a message arrives, with drag, resize and custom pet packs.

How to use it

1. Go to catpaw.meituan.com and download the installer for your system; on macOS, open the .dmg and drag the app into Applications.
2. On first launch click start to begin initial setup and pick an interface theme.
3. If you have used another IDE before, import its configuration so settings and habits carry over.
4. Register with a phone number or email; a new account starts with 500 free conversations.
5. Check credit usage in settings and request a top-up when needed; switch sandbox mode for tasks that need tighter isolation.
6. To cut repetition, turn frequent prompts into slash commands or record a browser flow into a reusable skill.

Pricing and requirements

• Trial: 0 per month with 200 credits each month and multi-agent calls included, meant for a first taste of the agent workflow.
• Pro: 88 CNY per month with 4,000 credits plus a 200-credit bonus, aimed at developers and heavy productivity users.
• macOS: arm build for Apple M-series chips (M1 to M4) and x64 for Intel, both requiring macOS 13 or newer.
• Windows: the x64 build covers Intel and AMD 64-bit processors on Windows 10 or newer; arm chips such as Snapdragon Windows devices are not supported yet.
• Mobile covers Android, iOS and iPad; Cloud mode needs no installation and runs in a browser.

Who it suits

• Developers who need a delivered result rather than a suggestion: describe a scrape-and-summarise job and get the table back.
• Front-end work, where the feedback loop for styling and interaction shrinks to a single conversation.
• Operations and store managers in local services, who get ready-made experts for store operations, reviews and marketing conversion.
• Teams that want their experience captured: recorded browser flows turn one person's method into a reusable asset.
• Teams already working in IM: digital employees inside Feishu or WeCom can be assigned work with an @ mention.
• People learning to code, since completion, explanation and examples sit in the same window.

Limits and caveats

• The Windows client does not support arm chips yet, so Snapdragon-based machines have to use Cloud mode.
• Usage is metered in credits: 200 per month on the trial tier, and multi-agent jobs consume far more than ordinary Q&A, so estimate before going heavy.
• The underlying models are listed as undisclosed (in-house plus external model capability), so which model performs which step cannot be verified from outside.
• Parallel agents and cloud long tasks bring convenience but also higher resource use and longer waits; a vague task description multiplies the rework.
• Full-access sandbox mode is broader, and high-risk operations require confirmation. For customer data, production systems and keys, start from workspace-write.
• The source does not state where data is stored, how long it is retained, or whether it is used for training, so confirm with the vendor before handling contracts, personal data or internal code.
• Connecting digital employees to Feishu or WeCom needs administrator rights in those tools, usually arranged by a company admin.
• The competitor comparison and ecosystem commentary come from a single third-party introduction page and should be read as reference, not as an independent benchmark.

Frequently asked questions

Is CatPaw a coding tool or an agent platform?

Both. It offers completion, ask-to-code and preview debugging on the IDE side, plus multi-agent collaboration, cloud long tasks and enterprise hosting; the official positioning is a full-scenario AI agent platform.

Is the free allowance enough?

A new account gets 500 conversations, and the trial tier adds 200 credits per month. That is workable for everyday Q&A; for complex multi-agent jobs, judge by actual consumption.

Does closing the laptop interrupt a task?

In Cloud mode tasks run 24/7 in the cloud and the vendor states local shutdown or disconnection does not affect execution; progress can be checked from other devices.

Which programming languages are supported?

The vendor lists mainstream ecosystems including Python, C++, Java, JavaScript, TypeScript, Go and Rust.

How is data protected?

The vendor cites two-level sandboxing, tenant isolation, centralised credential custody and tiered permissions, plus private deployment. Retention and training use are not disclosed in the source and need confirmation from the vendor.