Build an interface where agents and humans work with the same tools
Just as coding agents handle files, tests, and previews together, knowledge work needs an environment where you can see and modify an agent's process in real time. Agent Native is a framework designed so that the tools an agent uses and the buttons a user click share the same code. This ensures that what an agent does is immediately reflected on the screen, and any changes a user makes are picked up by the agent for the next step.
One action, multiple paths
The core of Agent Native is the concept of an 'action'. When a developer defines a specific function as an action, the agent calls it as a tool, and the UI executes the same function via code. In this process, validation, permissions, and implementation logic are all applied identically. Because the agent works directly through the same action layer as the UI, rather than clicking the UI with a mouse, data inconsistencies do not occur. For example, if you create an action to send a greeting, it returns the same result whether the agent uses it as a tool, a React component calls it, or an HTTP request is made.
A collaborative environment
This framework helps agents and users share the same data and state. Work processed by the agent is displayed in the UI immediately, and changes made by the user in the UI can be utilized by the agent in subsequent tasks. Additionally, UI states such as the current page, selected records, and active view are passed to the agent, enabling context-aware decisions. This structure creates an environment where the agent works independently while naturally accepting user supervision and intervention.
Included features and getting started
Agent Native provides agent chat, authentication and permission management, skills and memory, automation, agent teams, and a PostgreSQL backend by default. You can use PGlite for local development and PostgreSQL for production, running on Nitro-compatible hosts. Getting started is simple. Running the command npx --yes @agent-native/core@latest create my-agent --standalone --template chat creates an agent project based on a chat template. Users directly select and connect their own LLM, SQL database, tools, and infrastructure, and retain full ownership of all assets created.
Practical use cases
The README includes various open-source agent examples, such as meeting notes, design generation, slide editing, data analysis, calendar management, email processing, asset management, content writing, and planning. These examples serve as reference material showing how the framework applies to specific business areas. Each agent defines actions specialized for that task, following a flow where users review and approve results in a familiar way.
Considerations before implementation
This framework is TypeScript-based and released under the MIT license. Since users must prepare their own LLM, database, and infrastructure, the ability to manage these external dependencies is required. It also requires a Nitro-compatible host and a PostgreSQL (or local PGlite) environment. The structure where the agent and UI share the same action layer ensures consistency, but permission and validation logic must be carefully configured when designing actions. If complex permission systems or large-scale data processing are needed, it is advisable to review the scope of the provided authentication and permission management features in advance.
