DEV Community

Pratyush Ranjan Roul
Pratyush Ranjan Roul

Posted on

Building Cometflow Do AI: Our Journey Toward a Powerful Free AI Coding Partner

Building Cometflow Do AI: Our Journey Toward a Powerful Free AI Coding Partner

What if developers could get a powerful AI coding partner without depending on expensive subscriptions or paid API keys?

That's the idea behind Cometflow Do AI — a project we are now working on.

Our goal isn't to build another chatbot that simply generates code snippets. We want to build something much more useful: an AI software engineering partner that can understand projects, work with multiple files, debug problems, execute development tasks, and help developers build real software.

And we want to make it accessible for free.

Why are we building a separate product?

Initially, we considered adding AI coding capabilities to our existing Android application, Cometflow Do.

However, after thinking about the actual development experience, we decided against combining them.

Cometflow Do is focused on utilities and everyday tasks. Cometflow Do AI has a different purpose.

Serious software development often requires:

  • Working with complete project directories.
  • Navigating large codebases.
  • Editing multiple files.
  • Using a terminal.
  • Running tests and debugging errors.
  • Managing Git changes.
  • Reviewing and improving existing code.

A mobile application is not the ideal environment for all these workflows.

So we decided to keep the two products separate and make Cometflow Do AI a Windows-first desktop application.

This gives us more room to build a proper development environment instead of squeezing complex coding workflows into a mobile interface.

What are we trying to build?

Cometflow Do AI is being designed around a simple idea:

Give it a development task, and let it help you take that task from understanding to implementation.

Our planned capabilities include:

  1. Repository understanding

The assistant should be able to inspect an existing project, understand its structure, identify important files, and use relevant project context before making changes.

  1. Multi-file code generation

Instead of generating isolated snippets, it should eventually be able to create and modify multiple connected files while maintaining consistency across the project.

  1. Debugging and error fixing

We want it to investigate errors, understand their causes, propose fixes, apply changes with permission, and verify whether the solution works.

  1. Terminal and tool execution

A coding assistant becomes much more useful when it can interact with development tools.

Our planned architecture includes controlled terminal execution, project tools, testing, and Git integration — with permissions and safety controls rather than unrestricted command execution.

  1. Testing and self-correction

Writing code is only one part of software development.

We want the assistant to run available tests, inspect failures, and attempt corrections based on actual execution results.

The model strategy

We are researching a model-based approach built around Qwen3-Coder-Next.

Our current direction is to use the available instruction-tuned model as an initial working baseline while exploring specialised training and improvement using the base model.

We are also interested in:

  • Coding-focused fine-tuning.
  • High-quality training datasets.
  • Multi-teacher knowledge distillation.
  • Tool-use and agentic training.
  • Evaluation using real software engineering tasks.

However, we understand that a capable coding model is not enough by itself.

The surrounding agent architecture — repository indexing, planning, file operations, terminal tools, testing, and recovery — matters just as much.

Free access is a major goal

We don't want Cometflow Do AI to become another product where users immediately need a paid API key just to try its core features.

Our goal is to develop a model and architecture that can support accessible inference options, including local model execution where hardware allows and cloud inference where suitable infrastructure is available.

We also want to be transparent about limitations. Running powerful models requires computing resources, so providing unlimited cloud inference for free is a real engineering and infrastructure challenge.

We would rather build something sustainable than make unrealistic promises.

Why Windows first?

We are starting with Windows because it provides a much better environment for the workflows we want to support.

Our first milestone is a polished desktop application with a clean, natural interface — not a robotic-looking AI dashboard.

After building and testing the desktop version, we want to evaluate its actual performance, reliability, and usefulness.

Mobile development can be considered later, once the core product has been validated.

Our long-term ambition

We are not claiming that Cometflow Do AI is already one of the best coding assistants.

That's what we want to work toward.

Our long-term ambition is to make it a genuinely capable, free AI coding partner that developers can use for real software engineering work — from small fixes to complex projects.

We want to measure progress through real-world coding tasks, test results, reliability, and developer feedback rather than impressive marketing claims.

This is a long journey, and we are just getting started.

Cometflow Do AI — Building toward a better, more accessible AI coding experience.

Top comments (0)