Underthesea AI Assistant: Agentic Toolkit & Vietnamese NLP

What changed

The Underthesea AI Assistant, developed by undertheseanlp, has seen a recent release, version v9.5.0, marking a significant evolution into an open-source Agentic AI Toolkit. This update introduces built-in support for Vietnamese Natural Language Processing (NLP) alongside its existing multi-provider AI Agent capabilities. A key architectural change is the agent's communication mechanism, which now relies solely on Python's standard library (urllib and json) for interacting with LLM APIs, eliminating the need for external packages like openai, anthropic, or google-genai.

The toolkit supports a range of LLM providers, including OpenAI, Azure OpenAI, Anthropic Claude, and Google Gemini. Developers can configure these providers by setting environment variables or directly passing API keys and endpoints. The agent supports streaming responses and advanced features like tool calling, allowing agents to execute predefined functions. A comprehensive set of 12 default tools is provided, covering functionalities such as calculator, datetime, web search, Wikipedia lookup, file I/O, shell commands, and Python execution.

For managing complex workflows, Underthesea offers multi-session orchestration, enabling long-running agents with context reset and structured handoffs between sessions, inspired by Anthropic's harness patterns. Observability is enhanced through automatic tracing of agent calls to ~/.underthesea/traces/, with options for disabling traces or integrating with Langfuse for more advanced monitoring. The project also includes an optional agent server component (A2A protocol) that can expose agents over HTTP+SSE using JSON-RPC, with an optional bundled chat UI. This server component is a raw ASGI app, allowing integration with various ASGI servers without additional web framework dependencies.

Beyond agentic capabilities, Underthesea continues to provide a suite of Python modules for Vietnamese NLP. These include sentence segmentation, text normalization, diacritics restoration, word segmentation, part-of-speech tagging, chunking, named entity recognition, text classification, sentiment analysis, language detection, dependency parsing, translation, and text-to-speech.

Why it matters for builders

This update provides builders with a robust, open-source framework for developing AI agents that can interact with various LLM backends without vendor lock-in. The integrated Vietnamese NLP capabilities are particularly valuable for projects targeting Vietnamese-speaking users or requiring specialized language processing in that domain. The minimal external dependencies for agent communication simplify deployment and reduce potential conflicts.

Practical impact

Developers can leverage Underthesea to build sophisticated AI agents by easily integrating with their preferred LLM provider. The inclusion of default tools and the ability to define custom tools streamline the development of agents capable of performing complex tasks. For those working with Vietnamese text, the dedicated NLP modules offer a ready-to-use solution for tasks ranging from basic tokenization to sentiment analysis and NER. The agent server component also facilitates the deployment of agents as services accessible via standard web protocols.

Caveats and source limits

The provided source information details the features and architecture of the Underthesea AI Assistant. Specific performance benchmarks, pricing details for any associated services (though the toolkit itself is open-source), or detailed usage statistics beyond GitHub stars and forks are not present. The release date for v9.5.0 is not explicitly stated, though the excerpt mentions a "fresh release." The source also indicates that while the agent communication has zero external dependencies, the agent server component can optionally be installed with extras like uvicorn, starlette, and httpx.