Say it while it’s fresh.
Turn spoken ideas into an editable draft on your iPhone. Review your words before sending. Speech models are downloaded separately.
一呼YIHU
Step away from your desk. Keep your Agent work moving with your voice.
Illustrative scene and interface
Yihu is a native iOS browser shell optimized for Agent workbenches: open your own HTTPS web UI, transcribe speech drafts on the phone, retain voice notes, filter PM Radar items and stage content through system sharing. Apple Watch can receive paired-phone summaries. Yihu supplies no Agent backend; speech needs separately downloaded models and native enhancements depend on service compatibility. Merged source does not certify every device path.
Your familiar workbench. A native touch.
On-device speech / Voice notes / SharingTry the handoff. Play a short sample, edit the words, then place them in a workbench draft.
Your reviewed words will appear here.
You decide what to send, and when.
Turn spoken ideas into an editable draft on your iPhone. Review your words before sending. Speech models are downloaded separately.
Record now, transcribe in segments on your phone, and return to your audio and text later.
Bring text, links and files into the share inbox. Choose what to hand off to your workbench. Nothing is sent automatically.
Connect your own trusted HTTPS workbench. Yihu adds native iPhone capabilities around it.
Collie, OpenClaw and Hermes are examples of web workbenches you can open. Page access and native integration are separate capabilities.
Notifications and Live Activities require a compatible service. Yihu does not host Agents or provide a backend.
See what you needSwipe the table to compare prerequisites. With a keyboard, focus the table and use the arrow keys.
| Capability | What it does | Prerequisites and boundaries |
|---|---|---|
| On-device speech drafts | Transcribe on the phone and explicitly insert into a compatible draft; never presses Send. | Download compatible models and grant microphone access; host refusal or unknown editors can prevent insertion. |
| Segmented voice notes | Retain local audio and transcripts for later review. | Requires compatible models and microphone access; remaining recording paths need device acceptance, with no accuracy guarantee. |
| PM Radar task filters | Filter tasks per workbench and distinguish source-read time from offline cache. | Supply a JSON export or compatible HTTPS JSON feed; web cookies are not reused. Replies need separate authorization, phone confirmation and receipt readback. |
| Apple Watch summaries | Receive pending decision/acceptance summaries from the paired phone and hand off to it. | Requires watchOS 11+ and a paired phone; no standalone approval or direct Watch writes. Remaining device checks are pending. |
| System share inbox | Stage text, links and files before a user-selected handoff. | No automatic upload or send; the destination service's data practices still apply. |
Yihu is a native iOS browser shell around WebKit for your own reachable, trusted HTTPS workbench. Collie, OpenClaw and Hermes are web-access examples, not a certified native compatibility list. Yihu does not host Agents and is not their official client.
You need iOS 18+, your own trusted HTTPS workbench and, for the current source release, a Mac, Xcode, XcodeGen and your own signing setup. Speech also requires separately downloaded compatible models and microphone permission. Login and page compatibility depend on the server.
Yihu transcribes on the phone with separately downloaded Qwen3 ASR models. After review, text can be inserted into a compatible draft. Insertion respects explicit host refusal, does not force text into unknown editors and never presses Send. Incompatible pages retain text for explicit copying.
Yihu source includes longer recordings, segmented phone transcription and local audio/text retention, requiring microphone permission and compatible models. Remaining recording-related device checks are pending; no arbitrary duration, device performance or transcription accuracy is guaranteed.
Yihu PM Radar accepts your own JSON export or compatible HTTPS JSON feed, with per-workbench filters, management counts and bounded offline snapshots. Source-read times distinguish cached from fresh data; reads do not change ledger status. Authenticated feeds currently require a file export; Radar does not reuse web cookies or store credentials.
No. For eligible decision/acceptance items, Yihu requires separately authorized server writes and explicit phone confirmation, followed by receipt readback. It does not automatically change status, execute tasks, merge or deploy. Real replies still require separate device/service acceptance.
The watchOS 11+ companion receives pending decision/acceptance summaries through the paired phone and can hand off selected items or protected drafts to it. It does not provide standalone approval or direct Watch writes; the phone still requires confirmation and remaining pairing/device checks are pending.
Yihu's share extension stages text, links and files in an App Group inbox. It does not automatically upload or send them. Handoff requires a separate user action, and the destination workbench's data practices still apply.
No. Compatible downloaded models transcribe on the phone and voice notes are stored locally, but model downloads, web access, user sends/uploads and optional notification/relay integrations involve their respective services. On-device transcription is not a promise that all data stays on the phone.
No. Native notifications need a compatible web bridge, server endpoints and APNs configuration matching the signing environment. Live Activities also need a compatible relay you control. The placeholder relay is unusable, delivery needs separate acceptance and a ledger GET feed is not itself a push event source.
Yihu currently provides source code, not an App Store binary. Follow the repository build guide to configure your own identifiers, App Group and signing, then build for your iPhone. No preconfigured server or model files are supplied.
No. Yihu speaker verification is a convenience filter that can misidentify speakers, not authentication. Compatible WeSpeaker/CAM++ artifacts are not included; verification-dependent paths are unavailable without them. Verification and server write authorization must not be bypassed.
Open source. Built for iPhone. Connect your own services.
Currently a source release. Requires iOS 18+, your own workbench, speech models and signing setup.