How to Build an Agent Workspace on QwenPaw with Custom Skills and Streaming API

QwenPaw — is an instrumental environment for developing and testing AI-agents with support for pluggable model providers, user custom skills, and a local knowledge base. A new tutorial shows how to deploy a full-fledged workspace directly in Google Colab.

Installation and Project Structure

The process begins with installing the package via pip and initializing a working directory with the `qwenpaw init` command. The framework automatically creates a structured folder tree: separate directories for skills, knowledge files, and configuration. This allows you to immediately understand project organization and not waste time inventing structure from scratch.

After initialization, the tool generates a basic `config.yaml` configuration that the developer adapts to the specific project. The configuration describes model providers, agent parameters, and paths to additional resources.

Authentication is configured through environment variables. For local development, a `.env`-file is sufficient, but the tutorial recommends using the built-in Colab Secrets manager — this way API keys don't get into the code and don't leak when publishing the notebook.

Connecting Model Providers

One of QwenPaw's strengths is support for multiple language model providers without changing the core agent code. The tutorial demonstrates connecting the following providers:

• OpenAI (GPT-4o, GPT-4-turbo, and others)
• Anthropic (Claude 3.5 Sonnet, Claude 4)
• Alibaba Qwen (via Dashscope API)
• Ollama (local open-source models without API costs)
• Any OpenAI-compatible endpoint

Each provider is configured with a separate block in `config.yaml`. If desired, the agent can automatically switch between providers depending on the type of task or budget constraints.

Custom Skills and Knowledge Base

A key advantage of QwenPaw is the modular skill system. Each skill is a separate Python file with a natural language description of the function, a JSON schema of input parameters, and execution logic. The agent automatically selects the needed skill through tool calls. The tutorial shows creating several examples: web search, working with the file system, and custom computations. Each skill is tested in isolation before connecting to the agent — this significantly simplifies debugging and error localization. Beyond skills, the workspace can contain local knowledge files: markdown documents, JSON schemas, CSV tables. The agent accesses them through a built-in RAG mechanism without additional configuration, making QwenPaw convenient for creating specialized assistants for specific domains.

"The goal of

QwenPaw is to give the developer a ready-made scaffold where each component of the agent is isolated and easily tested separately," — the project documentation states.

Console and Streaming Testing

After setting up the workspace, the tutorial shows how to launch the agent's built-in console. It accepts text requests, passes them to the agent, and returns a response in streaming mode — text appears as it is generated, like in ChatGPT or Claude interfaces. Streaming mode is particularly useful for debugging: the console displays intermediate steps of agent work — skill selection, call arguments, intermediate results, and the final answer. This speeds up finding problems in agent logic compared to checking the final output. Besides console mode, QwenPaw provides REST API endpoints for integration with third-party applications. Both synchronous request-response and Server-Sent Events for real-time streaming are supported.

What It Means

QwenPaw solves a common problem in agent development: too much time goes into infrastructure instead of application logic. A ready-made scaffold with support for multiple providers, streaming, modular skills, and built-in RAG allows you to move to testing ideas in hours rather than days. The Google Colab format minimizes the barrier to entry — no local installations and environment setup required.