sptflo-local

Local-only swarm orchestration for SPT --- cloud-free agent coordination on Apple Silicon

Overview

sptflo-local provides on-device swarm orchestration without any cloud LLM dependency. It runs agent swarms entirely locally using Apple Foundation Models (Tier 1) and sptllm (Tier 2) for inference, coordinated over SPT Wire (QUIC-based transport) with SRP capability-based security. This crate is in early development.

Key Features

• Local-Only Execution --- all inference and coordination happens on-device
• Tiered Model Routing --- routes tasks to Foundation Models or sptllm based on complexity
• Agent Runtime --- lifecycle management for local agent instances
• Learning Loop --- adaptive learning from swarm execution outcomes via SONA
• Model Pool --- manages loaded model instances for concurrent agent inference

Dependencies

Key internal dependencies:

• spt-core (storage, hnsw) --- vector database for agent memory
• sptllm (candle, gguf-mmap) --- local LLM inference engine
• spt-sona --- adaptive learning
• spt-srp --- capability-based security protocol
• sptos-nucleus --- SPTOS sandbox primitives

Key external dependencies:

• spt-wire --- SPT Wire QUIC transport for agent coordination
• tokio --- async runtime
• dashmap --- concurrent state management

Usage

use sptflo_local::local_swarm::LocalSwarm;
// LocalSwarm::new(config).await to start a local swarm
// Agents are spawned and coordinated entirely on-device

API Reference

Modules

Module	Description
local_swarm	Top-level swarm orchestrator
model_pool	Manages loaded LLM model instances
agent_runtime	Agent lifecycle (spawn, execute, dissolve)
tier_router	Routes tasks to appropriate model tier
learning_loop	Adaptive learning from execution outcomes

Core Types

Type	Description
LocalSwarmError	Crate error type

Note: This crate is in early development. Module APIs are expected to evolve.