[ancilo]

With Ancilo, I wanted to try two things: make local AI accessible to people without a technical background, and develop an app in Rust even though I do not know the language myself. The result is an application for chat, documents, tasks and coding that runs AI models on the user's own computer.

Making local AI easier to set up

The applications I had tried felt cumbersome to varying degrees. Someone familiar with computers and programming can work through such a setup, but I wanted to build an app for people who simply want to try local AI without first having to understand its technical environment.

Ancilo therefore guides users through choosing and downloading a model. Setup takes the computer into account and offers controls for how much of its resources the AI may use. Bringing conversations, documents and tasks into one application is intended to make everyday use easier as well as the initial setup.

Hardware remains a constraint: available memory and processing power affect which models run usefully. Local processing keeps the content involved on the computer, while optional web search and cloud models involve connections to external services. The product website explains these distinctions and the requirements in more detail.

Developing in Rust with two coding agents

The second question concerned my own way of working. I wanted to find out whether coding agents could help me build an app in a language I do not know. I developed the foundations with Claude Code. During implementation, Claude Code repeatedly consulted Codex to check decisions, obtain reviews and work through suggested changes.

This back-and-forth was part of development from the beginning. In my judgement, it brought the quality to a high level early on. The app works for me, and I am delighted with the results so far. That is my experience of this project; it does not establish a general claim about the reliability of AI-generated code.

Existing implementations as references

I checked out the OpenCode and Codex codebases in directories alongside the Ancilo project. The agents could study approaches used by existing coding and chat agents and draw on them when developing the implementation. They did not have to devise every basic mechanism from scratch.

Ancilo can now also accept bounded tasks from Claude Code or Codex and carry them out with a local model. The calling agent receives the result for further processing. Running that task locally does not make the external agent's entire workflow local.

The Ancilo website covers features, requirements and getting started. The source code on GitHub is available under Apache 2.0; contributions, bug reports and specific feature requests are welcome.