Project Info

Flursor

This project did not submit a demo video on Devpost.

Inspiration

The project was inspired by Cursor. It is similar to Cursor, but for audio engineering or music production.

What it does

It has a promptable AI that integrates completely with natural-language prompts into the audio editing software REAPER. We chose REAPER since it was scriptable with Lua and inexpensive.

How we built it

The project has many components. It runs an MCP (model context protocol) server which integrates with REAPER directly using the OSC protocol and Lua scripts that can control REAPER. The MCP allows for an API from which an agent can make tool calls in REAPER. Then, we have our agent, which consists of a planner and executor and uses GPT-5. The planner plans out the tool calls to make based on the prompt, and the executor calls the MCP server to make said changes. Finally, we use MOSNet (MOS stands for Median Opinion Score) to gauge how the AI-determined MOS, or quality, of the audio changed after or before a prompt based on a 4-bar snippet. The MOSNet component is completely working and at the end of a prompt all data is sent to an endpoint to be cached for future training (MOS improvement, prompt, tool calls). However, the MOS integration with the agent fails sometimes when export of the audio from REAPER fails, which is something that can be improved on in the Lua script.

Challenges we ran into

We ran into many challenges, such as the MCP not working due to our Lua code being buggy, and our executor and planner having difficulty coming up with the current commands to the MCP server. In addition, we had to run all the different components together and fix CORS issues with requests.

Accomplishments we're proud of

We're proud of overcoming almost all our issues and coming to a robust MVP for agentic audio engineering, and there's still a long way to go for a full-fledged suite. However, we managed to get working Lua scripting and a completely working agentic AI system that was able to make tool calls in REAPER, building the MCP from scratch ourselves.

What we learned

We learned that multi-component projects like this can be super rewarding at the end when all the components work together and pay off for a good product.

What's next

We will have to fix issues in Lua with audio export and increase the functionality of the tool calls to making very specific selections and therefore be able to completely process very complex prompts, as well as incorporating reinforcement learning to train a model better to use our MCP.

Analysis

Compare with all teams

View

Metric

Abhibob
1

Figures cover GitHub contributors during the hackathon window. A co-authored commit counts in full for each author, so per-member totals add up to more than the whole-team figures.

Technology

Found in codeClaimed only

CSSIn code
FastAPIIn code
HTMLIn code
JavaScriptIn code
PythonIn code
ReactIn code
TensorFlowIn code
OpenAIClaimed

7 of 8 appear in the indexed code. 1 claimed on Devpost could not be matched to code, which may simply mean the tool leaves no trace in the repository.

AI coding agents

No AI coding agent signals were found in this repository.

Detected from committed agent config files and commit authorship. Absence of a signal is not proof an agent was unused.

Codebase size

Source size

68 KB

Source files

Counts recognized source files only; vendored directories, binaries and lockfiles are excluded, so this is smaller than the repository on disk.