Original Summary

I've been working on CoHamster, a fork of Cotabby that adds autocomplete to everyday typing on a Mac. Credit to FuJacob and the Cotabby contributors for the original app. This is my independent fork, and it stays free and open source under AGPLv3. The idea is simple: start typing in a supported text field, get a suggestion beside the cursor, and press Tab to accept it a word at a time. Or just keep typing. It's useful for emails, messages, notes, and other repetitive writing because you can finish a thought without opening a separate AI chat or copying text back and forth. Most of my changes are about making that interaction less frustrating: • Better word completion. It can finish a word you've only partly typed without repeating the beginning or treating it as a typo too early. If the model doesn't produce a usable ending, it can fall back to local dictionaries and words in your document or glossary. • Suggestions that keep up with you. Matching letters, spaces, and repeated Tab presses carry on through the existing suggestion. Local engines can also prepare the next few words while you're finishing a word or accepting a correction. You can choose suggestions mid-word or at word boundaries, and show one word or several. • Fewer interruptions. More stable ghost text, fixes for stale results after editing or switching fields, and short-lived memory of suggestions you've dismissed with Escape so they don't immediately come back. • Fresher screen context. Optional on-device screenshot/OCR context refreshes while you stay in the same field, keeps relevant excerpts, and clears stale context when a capture fails. Screen Recording is optional. • Work under the hood. Better reuse of the local model's prompt cache, revised context formatting and generation settings, and a 1,337-scenario benchmark plus typing and regression tests to check changes. I've also added support for text fields in the Codex desktop app. • A separate release you can actually install. CoHamster has its own hamster artwork, settings and model storage separate from upstream Cotabby, and a signed, Apple-notarized Mac build. Updates are manual through this fork's releases. Cotabby's emoji completion, inline math/conversion macros, and one-key autocorrect are still there too. You can run it entirely on your Mac using Apple Intelligence or a downloaded local model. No account, subscription, analytics, or telemetry. There's also an optional server connection inherited from Cotabby: it only sends text to a server if you explicitly choose and configure that engine. Stick with the on-device engines to keep generation all local. The current 0.6.3 build is a prerelease for Apple Silicon Macs on macOS 14+. Apple Intelligence needs macOS 26 and supported hardware. App compatibility and suggestion quality still vary, so I'd really appreciate reports of fields where it doesn't work or suggestions that get in your way. Download: https://github

  • 情报分类：技术学习与提效
  • 分类依据：内容涉及技术、AI、软件工具或工程实践
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/9/25 02:56:17