What I build on my own time.

After twenty years of programming, I wanted to build games, for my son and for myself, without hiring a team or giving one idea ten years. Swallowtail is the toolkit I built to make progress with the time I have.

Swallowtail lets me work with Godot from the terminal. I can inspect a scene, change it, run the game, and check what happened. Those commands can be saved in scripts or used by an AI tool while I review the changes. They also work without AI.

• A playtest can record frame times, inputs, checkpoints, and errors. Reports let me compare two runs.
• Release checks save logs, screenshots, and frame samples for comparison with earlier runs, then produce a PDF report.
• Adding a script to the game makes its commands available from the terminal and to connected AI tools.

I wrote HyperChi to reuse the common parts of a web application. It runs the application logic on the server. When someone submits a form or clicks a button, the server sends back the HTML that needs to change instead of reloading the whole page.

It includes form validation, sign-in, default security settings, and an embedded database. Utsushi, below, is built on it.

• Live updates for progress bars, shared boards, and chat.
• Back, forward and bookmarks keep working while parts of the page change.
• The documentation includes working examples and a version written for AI tools.

With Utsushi, I can take an existing clip and change who or what appears in it. Reference photos provide a replacement for a person; a written description specifies an object to change. The original camera movement, motion, and timing are preserved. It runs through ComfyUI on my own computer.

The open models run locally on a graphics card. Paid video services become available when an API key is added, with an estimated cost before each run. If Utsushi cannot estimate the cost, it asks for confirmation. I built it on HyperChi and the Comfy API Proxy.

• Video tools extract depth, pose, outlines, and cut-outs. They can also upscale a clip, smooth its motion, or extract frames and audio.
• Finished videos are saved to a local library, where a slider compares the original and edited versions.
• Progress is shown during each run. A side-by-side viewer lets you compare the result with the source.

Software written for Comfy Cloud expects its API. I wrote this proxy to provide the same interface for ComfyUI running on my own machine. Changing the server address lets the same software run jobs on a local graphics card.

It’s a Go version of Comfy’s Python proxy and passes all 82 tests in the original suite. In the published benchmarks, it handled 6 to 12 times as many requests per second. Memory use for large uploads was 27 MB, compared with a peak of 233 MB for the Python proxy.

• The proxy is a single download for Linux, macOS, or Windows.
• The included Go library can submit jobs to local ComfyUI or Comfy Cloud and save the results.
• Access is limited to the local computer until a token is configured. Requests from other websites are refused.