christopherkarani/Swarm

Type-safe tools, on-device inference, multi-agent workflows, memory, and guardrails, in one Swift-native runtime

What it solves

Swarm is a Swift-native agent runtime designed to help developers build production-ready AI agents for iOS, macOS, and Linux. It addresses the complexity of managing agent loops, tool integration, and model deployment by providing a type-safe, concurrent framework that supports both on-device and cloud-based inference.

How it works

Swarm uses a pluggable InferenceProvider system, allowing the same agent logic to work across different backends such as Apple Foundation Models (for on-device AI), Ollama, LM Studio, or OpenAI-compatible APIs. It leverages Swift 6.2's strict concurrency for safety and uses macros (via the @Tool macro) to automatically generate tool schemas from Swift types at compile time.

Who it’s for

Swift developers building AI agents for Apple platforms or Linux who need a robust, type-safe framework that supports on-device execution and complex agent orchestration.

Highlights

  • Composable Workflows: Chain, parallelize, or route agent tasks using Workflow and Job tools.
  • On-Device Inference: Native support for Apple Foundation Models for privacy and offline capability.
  • Resilience & Guardrails: Built-in support for retries, circuit breakers, rate limits, and input/output validation.
  • Durable Execution: Ability to checkpoint workflow progress and resume after a process restart.
  • Extensibility: Integration with Model Context Protocol (MCP) for tool discovery and OpenTelemetry for observability.

Related

  • Project
    eastriverlee/LLM.swift

    A Swift library for running local LLMs on Apple platforms, providing high-level abstractions for structured output, tool use, and chain-of-thought reasoning.

  • Project
    scouzi1966/maclocal-api

    AFM is a Swift‑based, OpenAI‑compatible server/CLI that runs Apple‑silicon‑native LLMs (Apple Foundation Models, Hugging Face MLX, DwarfStar) locally, offering chat, tool‑calling, vision, speech, and image generation APIs, plus a rich terminal UI and Swift packages for embedding.

  • Project
    AgentiLoop/Agent

    AgentiLoop Agent is a native macOS app that lets any LLM act autonomously on your computer—coding, building, automating apps, running shell commands, fetching the web, sending iMessages, and responding to voice. It supports 23 providers, offers a self‑verifying task loop, sub‑agents, custom MCP tools, and a Swift‑based scripting system, all distributed as a signed, notarized binary.

  • Project
    SharpAI/SwiftLM

    A native Swift, OpenAI-compatible inference server for MLX models on Apple Silicon, featuring TurboQuant (4× KV-cache compression), MTP speculative decoding, and SSD streaming for large MoE models, with extensive benchmarks on M-series chips.

  • Project
    huggingface/AnyLanguageModel

    A Swift package that provides a unified, drop-in replacement for Apple's Foundation Models framework, enabling support for multiple on-device and cloud LLM providers.