In this post Graham Helton tries out Crush, the AI coding agent that Charm released to run in the terminal, by using it to add Open Graph image generation to his personal static site — a tool that renders each post’s frontmatter into a press-release-styled HTML page and saves it as a PNG. He chose the task because he had done it before with other tools, it was low-risk, and the metadata it needed was already in each post’s YAML frontmatter. He reports finishing the feature and integrating it into his site’s build script with minimal hiccups.
The post is a firsthand account from one developer’s single project, and its conclusions are framed as his own: he liked the tool’s experience but would not use it for regular coding until he had self-hosted GPUs, because of what the API usage cost him.
Key Points
- The author found Crush to be what he expected from Charm: a TUI coding agent with Charm’s look and feel, which would feel familiar to someone who has used Claude Code.
- He ran Crush in a dedicated side pane inside neovim, which let him watch what it was doing and allow or deny its changes, or close it quickly while working manually; he reports some visual bugs when running it inside a neovim window.
- He found the diff UI jarring at first, though workable for prompts that make fairly precise edits.
- He singles out Crush’s list of changed files, its display of the model in use and the associated
cost, and its
ctrl-poptions — in particular quick switching between models and session summarization — as features he valued. - In his view the most important part of Crush, besides its TUI, is that it is model-agnostic, letting the user choose which models to use; he wants to try it with locally hosted models, which he says is not possible with tools like Claude Code.
- He reports that implementing the feature cost $23.04 in API fees, mostly using Sonnet 4 for the heavy lifting and Gemini Flash for quick edits and documentation updates, and took about 45 minutes against the few hours over a couple of days he estimates it would have taken him by hand.
- He argues that tools like Cursor have two pricing advantages over paying API fees directly: extensive optimization of what is sent to the model, including caching, and economies of scale from partnering with model providers — so that for about what this one feature cost him in API fees, a Cursor subscription would buy a very large number of prompts on larger models. This is his own assessment, supported by his own usage dashboard rather than any measurement.
- He still plans to use Crush for smaller tasks that do not need expensive models, such as answering questions about a fairly simple codebase or making quick changes to a homelab.
Context
The author writes that he wanted to try Crush because he dreads the experience of Cursor’s VS Code-centric UI, and that his site is for personal enjoyment rather than efficiency.