wechat-calls
Use wechat-calls for voice-session and media building blocks without the HTTP
daemon, local audio devices or OpenClaw. It includes the native direct/conference
session machinery used by the daemon, optional Opus, and portable PCM utilities.
All five public WeChat crates follow the same native release version; pin that
version when combining them.
Choose a feature set
| Cargo configuration | Available modules | Build requirements |
|---|---|---|
Default (native-runtime) |
PCM utilities, Opus, native session/media implementation and wechat-protocol |
Rust 1.93.1+, C/C++ compiler, CMake, Make and Perl |
default-features = false, features = ["native-opus"] |
PCM utilities and Opus, without protocol/account dependencies | Rust 1.93.1+, C/C++ compiler and CMake |
default-features = false |
PCM formats, queues, ordering, pacing, tones and mixing | Rust 1.93.1+; also builds for wasm32-unknown-unknown |
Native features are unavailable on WebAssembly. No feature opens a microphone or speaker, loads PulseAudio, or creates a Tokio runtime. Applications supply their execution, clocks and device adapters. The native media implementation can start workers when the embedding host requests it.
Mix PCM tracks
AudioFrame owns interleaved signed 16-bit samples. Its timestamp is a sample
position per channel on the shared presentation clock; generation fences
replaced audio, and sequence orders frames within a source. Announce tracks
before submitting frames. Render each due slot after the application's chosen
alignment delay; missing/muted sources contribute silence and late frames do
not accumulate a catch-up backlog. Do not submit your own uplink to the mixer.
use ;
let format = new?;
let mut mixer = new?;
mixer.add_track?;
mixer.push?;
let mixed = mixer.render?;
assert_eq!;
# Ok::
AudioQueue fences interrupted output with a new generation. PcmOutput admits
variable-size PCM chunks with backpressure and produces frames. FramePacer
avoids bursts after a scheduling stall. These are scheduling tools, not a full
jitter buffer or congestion controller. Clearing application audio never
licenses resetting a transport counter or reusing a key.
Check examples offline
From a source checkout:
# tracks=2 samples=320 first_sample=3000
# frames=50 packet_bytes=... decoded_samples=16000
OpusEncoder and OpusDecoder implement AudioEncoder and AudioDecoder for
one configured frame per raw Opus packet. Authenticate/admit packets before
decoding them. Neither example accesses accounts, devices or the network.
Native calling integration
wechat-calls depends on wechat-protocol; protocol does not depend on calls.
conference_media decodes admitted conference sources; conference_send handles
one paced uplink. The admission module provides the same per-account
controller/attempt ownership state machine used by Core; it performs no I/O and
holds no account credentials. Core retains daemon permissions and journals.
The admission, session and native_media modules are lower-level host
integration APIs hidden from generated docs while the public account-bound
CallClient/CallHandle facade is unfinished. There is no client.calls() API
in this release. Embedding those modules requires account ownership, permission
checks, request journals, cancellation, event delivery and uncertainty handling.
Do not retry signaling that may already have reached the server.
For an application that needs complete dial/join/invite/leave operations today, use the daemon's Calls API. Its JSON-v2/base64 PCM contract is separate from this crate's typed frames and WeChat media packets. Native workers currently stop at 20 minutes; key renewal, long-call acceptance, video and direct-to-conference promotion are not provided.
Our code is MIT-licensed; see LICENSE. Native Opus distributions must also retain NOTICE.md, LICENSE-OPUS and LICENSE-OPUSIC.