HuiZiGe
Let several models suggest your next move.
macOS 14+ · Apple Silicon / Intel · Primarily Chinese UI
App + source · Redeem once, download again any time
Bring your own model service. Setup guide on the download page.
v0.4.9 fixes a Codex issue where clicking a regular suggestion inserted the text but did not send it.
Latest GitHub release (repository access required)
GitHub installation and supported tools
Sign in to GitHub with repository access. Download HuiZiGe-macOS.zip and SHA256SUMS.txt and verify the ZIP using the release instructions. Unzip the app, move it to Applications, grant Accessibility access when prompted, and configure your model service.
Supported tools include Codex Desktop, Code sessions in Claude Desktop, Claude Manager, DeepSeek Harness (DSH), ZCode, Kimi Work, local Pi Web in Chrome or added to the Dock from Safari, pi in iTerm2, and Hermes Agent in iTerm2 or macOS Terminal with the bundled plug-in enabled.
HuiZiGe v0.4.7: Suggestions while a task runs · 1:07
v0.4.2 · Original product demo · 3:47
Demo notes and credits
Conversations and surrounding tools are demonstrations. Narration is AI-generated. “Neon” — Scott Buckley, source, CC BY 4.0; No Melody excerpt, equalized, faded and mixed. Xiaolai / 小赖字体 — SIL OFL 1.1, font source.
Community stats
Sample details All preferences
Finalized rounds with a displayed suggestion count as finalized displays above.
Stats generated (Beijing time). How is this calculated?
Models
Filter models and comparisons.
| Model / effort | Adoption rate | Devices | Selected / displayed | Sample |
|---|
Compare
Same prompt. Both available at selection.
Daily trend
Overall selection rate · All preferences
How are the stats calculated?
- Display and selection
- A suggestion counts as displayed when it is visible and selectable. A later user message finalizes the round. The last successful insertion counts as its selection and may be edited before sending. Pending and interrupted rounds are excluded from adoption rates.
- Adoption rate
- Selected rounds ÷ finalized displayed rounds, calculated per device and averaged equally. The pooled round rate in details combines all rounds, so it can differ. Device counts include devices with finalized displays; one person may use several devices.
- Comparisons
- Both candidates must be visible and selectable at the moment of choice, with the same prompt. Rates use rounds selecting A or B, averaged per device. Neither, other selections and identical text have separate counts.
- Scope and samples
- The overview and daily trend cover all preferences. Filters apply to models and comparisons. Dates use the round's start day in Beijing time; late finalization updates that day. Fewer than 3 devices or 30 finalized displays is a small sample. — means no valid rate; 0% means displayed but never selected.
- What the numbers show
- These are choices made during real use, influenced by preferences, prompts and arrival times. Effort labels come from request parameters. All supported tools contribute to model totals under metrics v1.
What is shared?
Sharing is off by default and covers only new rounds after opt-in. Reports include public model names, requested efforts, preference kinds, dates and counts, plus an anonymous device ID, public key and version for verification and deduplication. This page shows community aggregates.
Community reports exclude conversation, candidate and prompt text, accounts, session IDs, service names, addresses and credentials. Preview the report before opting in, then pause or delete server records at any time.
Contribute
Opt in via HuiZiGe → Settings → Adoption → Share with Radar Community.