Skip to content

About

An open interoperability layer between wearable devices and AI agents

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Repository files navigation

AgentWearLink

Swift Core CI Docs CI License: Apache-2.0

Connect wearable devices to existing AI agent runtimes—without coupling your application to a device vendor or model provider.

한국어 · Get started · Architecture · Documentation · Contributing

Important

Pre-alpha / integration preview. The Swift packages, vendor-backed simulator tests, and isolated real OpenClaw Gateway contracts are implemented and CI-tested. The reference Ray-Ban Meta + iPhone + Mac/Tailnet deployment has not been physically validated. This is not a production-readiness or hardware-compatibility claim.

Overview

AgentWearLink (AWL) is a Swift interoperability layer for wearable input, agent execution, and host-owned output. It translates device capabilities and events into vendor-neutral contracts, delegates intelligence to an existing agent runtime, and routes responses to a replaceable output sink.

flowchart LR
    subgraph Device["Device integrations"]
        M["Meta Wearables DAT"]
        F["Future wearable SDKs"]
    end
    subgraph AWL["AgentWearLink"]
        D["DeviceAdapter"]
        C["Core: capabilities, events, lifecycle"]
        A["AgentAdapter"]
        O["InteractionOutputSink"]
    end
    subgraph Runtime["External agent runtimes"]
        G["OpenClaw Gateway"]
        R["Future runtime adapters"]
    end
    M --> D
    F -.-> D
    D --> C
    C --> A
    A --> G
    A -.-> R
    A --> C
    C --> O
    O --> S["Host output: iOS TTS / UI"]
Loading

AWL does not replace an agent framework. Models, memory, tools, RAG, MCP, orchestration, and sessions remain the responsibility of the connected runtime.

Reference stack

Layer Current reference
Wearable Ray-Ban Meta via Meta Wearables Device Access Toolkit (DAT)
Host iPhone, iOS 17.2+ reference integration
Connectivity Private Tailscale deployment
Agent OpenClaw Gateway on a Mac (reference: Mac mini)
Output Host-owned Apple speech synthesis via AVSpeechSynthesizer
Vision Explicit, bounded, event-driven camera snapshots

These are replaceable reference integrations, not dependencies of AgentWearLinkCore. Wearable PCM capture and unattended background camera access are not implied by DAT speech or invocation support.

Quick start

For the vendor-neutral Swift package, use Swift 5.10+ and macOS 14+:

git clone https://github.com/JDeun/AgentWearLink.git
cd AgentWearLink
swift test

The concrete Meta DAT adapter has a separate Swift 6.0+ toolchain requirement to resolve the pinned vendor SDK; its integration sources currently use Swift 5 language mode. See the installation and verification guide for the iOS simulator and OpenClaw pathways.

Packages

Product Purpose
AgentWearLinkCore Vendor-neutral events, capabilities, interaction lifecycle, bounded streaming and vision contracts
AgentWearLinkMetaDAT Device adapter boundary; concrete vendor integration is in Adapters/MetaDAT
AgentWearLinkOpenClaw Gateway WebSocket, device identity, pairing, streaming, cancellation and reconnect
AgentWearLinkAppleOutput Optional host output sink and Apple TTS
awl-openclaw-probe Explicit read-only Gateway connectivity probe
awl-openclaw-chat-probe Explicit, potentially mutating agent turn probe

Validation status

Evidence level Status What it establishes
Swift Core and adapter tests CI passing Deterministic contracts and error handling
Meta DAT compile / MockDeviceKit CI covered Pinned SDK and simulator-hosted integration; not physical hardware
Real pinned OpenClaw Gateway CI passing Isolated Gateway auth, exact-ID approval, grant reuse/revocation, streaming, abort, and two-turn session smoke
iPhone ↔ Tailscale ↔ actual Mac Not yet accepted Requires private deployment testing
Physical Ray-Ban Meta camera/audio/voice Not yet accepted Requires wearable, iPhone and permissioned DAT app

The real Gateway checks use a disposable localhost Gateway and a synthetic local model. They do not validate personal sessions, real-model memory/tool behavior, a human approval ceremony, iOS Keychain entitlements, or wearable hardware. See evidence and CI gates, open acceptance work, and the validation matrix.

Engineering principles

  • Portable Core: Vendor SDK types and agent-specific protocols stay outside Core.
  • Least capability: Advertise only supported, currently available device actions.
  • Explicit capture: Camera snapshots are opt-in and bounded; continuous video is not the default.
  • Safe delivery: Cancellation is idempotent; reconnect must never silently replay an uncertain mutating request.
  • Data minimization: No private media retention by default; secrets stay in platform secure storage.
  • Evidence-based claims: Simulator, isolated integration, deployment, and physical acceptance remain distinct.

Roadmap

This is an acceptance roadmap, not a promise that hardware gates are already passed.

Phase Acceptance target Tracking
P0-A Real Ray-Ban Meta + iPhone DAT baseline #1
P0-B Actual iPhone → Tailnet → OpenClaw text round-trip #56
P0-C / P0-D Wearable audio output and hands-free voice invocation #5, #6
P1 Explicit photo → vision agent on physical devices #58
P2 Hardware reliability, privacy and recovery matrix #59
Later Additional SDK/runtime adapters driven by demonstrated need Issues

Documentation and support

Start with the documentation index or the getting-started guide. The architecture, OpenClaw integration, Meta DAT integration, and product requirements describe the contracts in more depth.

Bug reports and proposals are welcome through GitHub Issues. Before contributing, read CONTRIBUTING.md; report sensitive issues according to SECURITY.md.

License: Apache-2.0.

About

An open interoperability layer between wearable devices and AI agents

Resources

Contributing

Security policy

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages