Speak and listen.
The browser workshop now supports four Kokoro voices, read-aloud, WAV export and Whisper transcription. Audio goes into the composer for review. Physical microphone acceptance remains pending.
A personal AI.
A node we can build together.
Bring the latest local AI experience to Linux. Get the working text service alongside a developer handoff for speaking, listening, model controls and reviewed contribution exports. Voice cloning and image generation are documented as separate experiments.
Text service + current browser feature handoffText starter, browser reference source, model manifests and Linux acceptance checks in one ZIP. This handoff update does not add voice or images to the Linux service. Models download separately from wyck.ai, with exact sizes and SHA-256 checks. No account needed.
From download to your first reply.
Start with Linux x86-64, Node.js 22.x, and npm. The starter uses your CPU. Allow about 410 MB for models, plus dependencies. Hardware requirements and reply speed still need contributor measurements.
unzip wyck-linux-contributor-kit-0.1.2.zip cd wyck-linux-contributor-kit-0.1.2 npm ci npm test npm run models npm start
In another terminal in the same folder, run npm run chat. Wait for the memory check and test reply to succeed. Type /fact diagnostic_fee 89 dollars, then ask about your diagnostic fee. Use /ask for general model conversation without personal memory. /memories shows what you kept; /forget ID removes it.
The service stays on your machine. Memory is stored locally, and chat is not automatically saved or shared. Read the included guide before connecting to a remote server.
The local AI features to bring to Linux.
Manage local models.
Text, Speak and Listen have progress and Load/Stop controls. Optional packs have verified downloads, cancellation and removal. The handoff documents startup ordering and the Linux storage adapters needed.
Review and contribute.
Preview a practice result with its model version, timing and feedback before exporting it locally. Reference recordings, voice profiles and personal memory are excluded. Nothing is automatically uploaded.
Separate experiments: session-only voice cloning has functional browser evidence but still needs similarity acceptance. The current image prototype fails the clean-icon quality bar and is research-only. Neither is a released Linux feature.
Read the Linux developer handoff ↗
The ZIP includes exact browser reference files and a SHA-256 source inventory, pinned build recipes and four model-pack manifests. These are porting references, not a standalone desktop application. Speaking/listening model files add about 138 MB; runtimes are additional and optional models are not downloaded by the text starter.
Three ways to contribute.
Make Linux feel solid.
Use the handoff to port speaking/listening and model controls. Reproduce installation, measure real Linux performance, and validate capture, cancellation and private storage.
Connect two nodes.
After the Linux release gates pass, start with the signed test vector. Agree on identity, consent and message formats. Prove a direct connection, rejected tampering and recovery after a disconnect.
Improve the model.
Run the baseline evaluation. Compare memory accuracy, useful replies and hardware cost. Submit reproducible changes with consented data and honest results.
What works. What comes next.
This kit includes a CPU inference service, a terminal client, explicit local memory, verified downloads, synthetic evaluation probes, and signature test vectors. VALIDATION.md records executed checks and known limits. The text-only runtime removes Sharp from installation. Personal answers use explicitly saved facts with provenance; missing or conflicting facts receive a fixed refusal. General model replies remain unverified. The unchanged 0.1.1 text service passed Ubuntu 24.04 and Arch container checks, including real inference with external networking disabled. Those results do not validate a Linux voice or image port. View the Linux validation.
Peer discovery, attestation, shared-feed replication, browser memory interchange and a desktop application are not implemented in this preview. The ANT draft is a design input; it does not mean its endpoints are live.
Our first shared milestone: two Linux machines exchange one approved synthetic insight, reject tampering and replay, and resume after reconnecting. Personal memory stays private.
Bring evidence we can build on.
Include the kit version, model hashes, OS, hardware, reproduction steps and test results. Keep private memories and credentials out of reports. Model changes should include the baseline and held-out results, plus the training and export recipe when applicable.
The ZIP contains the contributor source. New starter code is Apache-2.0. Third-party model and dependency terms are documented inside the kit.