Feishu-Native AI Agent: Proma Brings Proactive Agents to Chinese Workplace

Full English Translation

Proma: The Smoothest Universal Agent Experience, Built for 100x Professional Users

Proma is an open-source project that brings proactive AI agents into Feishu (Lark) group chats. Built on the Claude Agent SDK, it demonstrates a complete reference implementation for embedding agent capabilities into Chinese workplace communication platforms.

What Makes Proma Different

Proactive Rather Than Reactive Traditional chatbots wait for a user to @mention them. Proma reads group chat context continuously and decides when to contribute — suggesting actions, surfacing relevant information, or offering to complete tasks. This “ambient agent” model means the agent becomes a persistent team member rather than an occasional tool.

Claude Agent SDK Foundation The project uses Anthropic’s Claude Agent SDK as its agent runtime, demonstrating how Western agent infrastructure can be adapted for Chinese deployment scenarios. The SDK provides the reasoning, tool-use, and multi-step execution capabilities, while Proma adds the Feishu integration layer.

Model-Agnostic Architecture While built on Claude Agent SDK, Proma supports flexible model provider configuration — users can switch between Claude, Qwen, DeepSeek, or any compatible model provider. This is critical for Chinese deployments where model availability and cost considerations differ from Western markets.

Native Feishu Integration The agent appears as a regular group chat member in Feishu:

• Reads and understands ongoing conversations
• Responds to natural language requests
• Proactively suggests actions based on context
• Can be configured with different personas and capabilities
• Supports multi-turn, multi-participant conversations

Proactive Agent Architecture

The proactive agent loop works as follows:

1. Context Collection: The agent continuously monitors group chat messages and builds an understanding of the ongoing discussion context.
2. Relevance Assessment: An internal scoring mechanism evaluates whether the agent should contribute based on the conversation topic, its configured capabilities, and conversation dynamics.
3. Action Generation: When the threshold is met, the agent formulates a response or action — this could be answering a question, suggesting a workflow, or offering to execute a task.
4. Execution: The agent uses its tool set to complete requested actions, reporting results back in the chat thread.
5. Learning: Over time, the agent adapts its proactivity threshold based on user feedback and interaction patterns.

Deployment Model

Proma is designed for self-hosting:

• Deploy the agent service on your own infrastructure
• Connect to Feishu via the official API
• Configure model providers through environment variables
• Manage agent behavior through configuration files

This self-hosted model addresses Chinese enterprise concerns about data security and infrastructure control while providing the convenience of a chat-native agent experience.

Significance for the Ecosystem

Proma represents a convergence of several important trends:

• Western agent infrastructure (Claude Agent SDK) being adapted for Chinese platforms
• The shift from reactive chatbots to proactive ambient agents
• Open-source implementations that enterprises can self-host
• Native integration with Chinese workplace communication tools

As Feishu, WeCom, and DingTalk compete for enterprise messaging dominance, the platform that best supports agent integration may gain a significant advantage.