Polymr – A Platform for Controllable Execution and Adaptive Interaction

Hey HN, I’m open-sourcing Polymr, an agent platform that focuses on controllable execution and adaptive interaction through dynamically generated tools and interfaces.

A core design principle is that MCP tools, generated scripts, frontend components, and pages all operate on the same underlying tool surface and can build on top of one another.

Control comes from things like:

- Standardized preview mode when a tool call is not automatically approved. I added an SQL MCP server that can run a DML in a transaction and roll it back to provide the user with an exact preview of the resulting changes.

- Granular permissions in a hierarchical model that balances power and ease of use. I added a CLI MCP server that introspects the script with Tree-sitter and dynamically extracts the required permissions and prompts the user only when additional permissions are needed.

- Runtime policies to restrict tool calls ad-hoc, for example restrict which folders can be accessed or make them immutable.

- Polymr has no "mandated" tools out of the box, allowing the user to shape it as needed.

The adaptive interaction includes:

- The LLM can create scripts that encapsulate domain-specific logic on top of existing MCP tools. These scripts can be reused by other scripts, frontend components, pages, or exposed back as MCP tools. For example, a generic SQL tool can be wrapped in a higher-level tool that implements business-specific object mutations while hiding the underlying queries.

- The LLM can create frontend components and pages that also work on the same tool surface, allowing for a mixture of traditional web applications combined with LLM interaction. Everything that is available on the screen is, by its very nature, also approachable by the LLM.

- The LLM can reuse existing components or create entirely new ones to create dynamic composite views on the fly in a session, tailored to the problem at hand.

While I use it daily for both work and outside of work, for coding and non-coding, it still remains very much a personal project, so expect to find bugs and rough UI edges here and there. Any feedback, critique, or bug reports are highly appreciated!