🌳 Orchard

A decentralized, privacy-preserving swarm-intelligence network powered entirely by idle Apple devices.

Orchard turns millions of idle iPhones, iPads, and Macs into a cooperative supercomputing network. By coordinating the on-device Apple Foundation Models (AFM) and Apple Neural Engines (ANE) of participating devices, Orchard tackles massive, complex problems — from disease modeling to decentralized web indexing — while keeping 100% of raw user data on the device.

---

Why Orchard

Conventional cloud AI	Orchard
Energy-intensive centralized datacenters	Idle silicon you already own
Per-token API fees	Zero marginal compute cost
Raw data leaves the device	Raw data never leaves the device
Owned by a single company	Collectively owned knowledge graph

The Four Technical Pillars

1. Local Node Execution — Each device is a node running native on-device inference via the Foundation Models framework on the Neural Engine. Opportunistic: runs only when plugged in, on Wi-Fi, and idle.
2. Local Micro-Swarms — Devices on the same LAN shard high-parameter models across their unified memories using MLX-style peer-to-peer clustering.
3. Global Agentic Workflows — A Task Router (on Private Cloud Compute) fragments massive problems into micro-prompts; millions of edge agents process chunks and return structured output for consensus aggregation.
4. Cryptographic Privacy — Federated Learning with Secure Aggregation (SecAgg) and Differential Privacy. Only masked gradient updates leave the device.

See DESIGN.md for the full architecture and docs/ for the public site.

Status

🌳 All five milestones (M1–M5) landed. OrchardNode runs a NodeRuntime actor that gates work behind the opportunistic scheduler and does structured-output inference via Apple's Foundation Models on OS 26+ (deterministic stub fallback elsewhere). OrchardSwarm is the micro-swarm coordination layer — peer discovery, coordinator election, and a memory-aware pipeline-parallel layer-shard planner. OrchardRouter is the global layer — job fragmentation, redundant load-balanced assignment, and majority-vote consensus that outvotes and flags faulty nodes. OrchardCrypto is the privacy layer — Bonawitz-style Secure Aggregation (Curve25519 + exact pairwise mask cancellation) and differential privacy, recovering the federated mean from masked vectors alone. OrchardPilot ties them together end-to-end on a real scientific workload. The micro-swarm ShardExecutor seam is implemented too: a pure-Swift PipelineRunner runs a ShardPlan as a distributed forward pass (bit-identical to monolithic), and OrchardMLX provides a real Metal-accelerated executor on mlx-swift (opt-in; just mlx-demo). OrchardTransport ships activations between pipeline stages over real Network.framework TCP, so PipelineRunner drives a genuine multi-device pipeline (verified over localhost: distributed == monolithic), and nodes find each other over the LAN via Bonjour auto-discovery (just bonjour-test). See the roadmap.

Quick start

just setup     # install toolchain + resolve packages
just build     # build all targets
just test      # run the test suite
just demo      # run one task through a node (Foundation Models on OS 26+, else stub)
just pilot     # run the full pipeline: distributed scan → consensus → federated refinement
just site      # preview the GitHub Pages site locally

just pilot drives a scientific workload through all five layers at once:

🌳 Orchard pilot — distributed folding scan
  scorer placement:  solo (scorer fits on one device)
  candidates:        12 evaluated across 4 nodes
  consensus:         12/12 (dissents rejected: 9)
  best conformation: [0.486, 0.336, 0.543]  energy 0.4708
  refined (federated DP gradient step):
                     [0.487, 0.208, 0.593]  energy 0.3009
  → energy reduced by 0.1699 toward the native state [0.500, -0.300, 0.800]

just demo on an OS 26+ machine produces a real on-device structured summary:

{
  "title": "Apple Neural Engine Accelerates On-Device ML",
  "summary": "The Apple Neural Engine enhances machine learning directly on devices…",
  "topics": ["Apple", "Neural Engine", "On-Device ML", "Machine Learning"]
}

Requires macOS 26+ / Xcode 26+ for the on-device Foundation Models APIs. See TOOLS.md.

Repository layout

orchard/
├── DESIGN.md                    # Architecture & system design
├── TOOLS.md                     # Toolchain, frameworks, dependencies
├── AGENTS.md                    # Guidance for AI coding agents
├── CLAUDE.md                    # Claude Code working agreement
├── justfile                     # Task runner
├── Package.swift                # SwiftPM manifest
├── Sources/
│   ├── OrchardProtocol/         # Wire types, task specs, schemas, node capabilities
│   ├── OrchardNode/             # Node runtime, scheduler, Foundation Models adapter
│   ├── OrchardSwarm/            # Peer discovery, coordinator election, layer-shard planner
│   ├── OrchardRouter/           # Job fragmentation, redundant assignment, consensus aggregation
│   ├── OrchardCrypto/           # Secure Aggregation (SecAgg) + differential privacy
│   ├── OrchardPilot/            # Capstone: one scientific workload through every layer
│   ├── OrchardMLX/              # Metal-accelerated ShardExecutor on mlx-swift (opt-in)
│   ├── OrchardTransport/        # Cross-device transport over TCP + Bonjour auto-discovery
│   ├── orchard-demo/            # Node + swarm + router + crypto + transport demo executable
│   ├── orchard-pilot/           # End-to-end folding-scan pilot executable
│   └── orchard-mlx-demo/        # MLX sharded-execution demo (opt-in, Metal)
├── Tests/                       # Swift Testing suites
├── docs/                        # GitHub Pages site
└── .github/                     # CI + Pages workflows

License

MIT — concept and scaffolding. Not affiliated with or endorsed by Apple Inc.