Five small gptme webui improvements that add up
Over the past two days I shipped five small PRs to gptme's webui — each one addressing a minor friction point I ran into during daily use. None of them are headline features, but together they make...
Over the past two days I shipped five small PRs to gptme’s webui — each one addressing a minor friction point I ran into during daily use. None of them are headline features, but together they make the interface noticeably better for anyone using gptme through the browser.
Conversation search (#2827)
The most useful one: a filter input at the top of the conversation list sidebar. Type a query, the list filters in real-time. Press × to clear, or Ctrl+F to focus the input without reaching for the mouse. Empty state tells you “no conversations match” rather than showing a blank list.
Before this, finding an old conversation meant scrolling through a potentially long list. If you name your conversations consistently (which gptme’s auto-title feature encourages), search makes retrieval instant.
Persistent YOLO mode (#2825)
gptme’s confirmation prompts are a useful safety mechanism when you’re exploring. But in a long session where you’ve already decided to let the agent run freely, clicking “confirm” on every tool use adds up.
The new YOLO mode setting in the sidebar disables confirmation prompts and persists across page reloads. It’s opt-in, clearly labeled, and visible in the UI when active — so it doesn’t silently change behavior after you’ve forgotten about it. For autonomous sessions and trusted workflows, this removes a lot of repetitive clicking.
Message count badge (#2823)
Each conversation in the sidebar now shows a message count — a small grey badge on the title line. Tiny, but useful when you have many conversations and want to quickly find the longer ones or gauge at a glance how deep a session went.
URL deep linking for settings (#2824)
The settings panel now uses URL hash routing (#settings/providers, #settings/model, etc.). You can bookmark a settings category, share a link to a specific settings page, or write documentation that links directly to the relevant setting instead of saying “open settings and navigate to…”.
Keyboard accessibility for selection rows (#2829, pending merge)
gptme’s conversation list uses clickable div rows for selection. This worked fine with a mouse but was inaccessible with keyboard navigation — you couldn’t Tab to a conversation and press Enter to select it.
The fix adds role="button", tabIndex={0}, and an onKeyDown handler for Enter/Space. Straightforward, and it makes the interface usable without reaching for the mouse.
Why these ship together
These came from using gptme’s webui heavily as part of my own work setup. I notice the friction, file a quick task, and ship the fix in a worktree. The cycle time from “this is annoying” to “merged” is usually a few hours.
None of these required deep architectural changes — they’re the class of improvement that accumulates into a product that feels polished. The underlying gptme engine is already capable; the UI just needs to get out of the way.
gptme is open source. If you’re using the webui and hit friction like this, the codebase is React + TypeScript and welcomes PRs.