Contents

christopherkarani/swarm

Install

Default link is lean: core Swarm + on-device Foundation Models. Graph/memory/web/Hive paths are trait-gated (off by default) and are not linked into Swarm unless you enable Integrations.

HiveCore, Membrane, and ContextCore are native in-tree Sources/ targets (internal modules — not separate library products). Enabling Integrations links those modules plus Wax (still remote) and SwiftSoup; omitting the trait does not link them into Swarm. Lean resolve never pulls Hive/Membrane/ContextCore/Conduit package identities, and (with trait-gated product edges) also does not pin Wax, MetalANNS→GRDB, swift-crypto, swift-mutex, or SwiftSoup. Always-on remotes remain (swift-syntax, swift-log, MCP sdk, OTel, plus NIO transitives — including swift-collections via NIO). SWARM_CORE_ONLY=1 drops the integration package block entirely. ContextCore / full Membrane session stack require Apple platforms (Metal/CoreML); Linux Integrations still builds Hive + MembraneCore + web helpers. DefaultAgentMemory uses a CoreML embedding model that is not bundled — without minilm-l6-v2.mlpackage, ContextCore falls back to deterministic pseudo-embeddings.

Root-package note: bare swift build / swift test on this repo compile every registered target, so integration modules need either --traits Integrations or the lean CI helper (scripts/ci/lean-build-test.sh). App consumers only build reachable targets and stay lean without that helper.

// Lean default link (recommended for most apps)
.package(url: "https://github.com/christopherkarani/Swarm.git", from: "0.6.0")

// Full integrations: durable Hive workflows, ContextCore+Wax default memory,
// Membrane adapters, and web helpers
.package(
    url: "https://github.com/christopherkarani/Swarm.git",
    from: "0.6.0",
    traits: ["Integrations"]
)

From a checkout of this package:


swift build --product Swarm --product SwarmMCP --product SwarmOpenTelemetry \
  --product SwarmMembrane --product SwarmCapabilityShowcase
# Full graph
swift build --traits Integrations
swift test --no-parallel --traits Integrations
swift run --traits Integrations SwarmCapabilityShowcase matrix

Quick Start

import Swarm

// The @Tool macro generates the JSON schema at compile time
@Tool("Looks up the current stock price")
struct PriceTool {
    @Parameter("Ticker symbol") var ticker: String
    func execute() async throws -> String { "182.50" }
}

// Create an agent with unlabeled instructions first and tools in the trailing @ToolBuilder closure
// Built-in backend: Apple Foundation Models (no API key on supported devices)
let agent = try Agent("Answer finance questions using real data.",
    configuration: .default.name("Analyst"),
    inferenceProvider: .foundationModels()) {
    PriceTool()
    CalculatorTool()
}

let result = try await agent.run("What is AAPL trading at?")
print(result.output) // "Apple (AAPL) is currently trading at $182.50."

That is a working agent with type-safe tool calling. Swarm also supports AGENTS.md and SKILL.md for declarative agent specs and reusable skills — see the Getting Started guide for the full workspace layout.

Why Swarm

  • Swift concurrency is part of the surface. Swift 6.2 StrictConcurrency is enabled across the package.
  • Tools stay type-safe. The @Tool macro generates JSON schemas from Swift structs.
  • Workflows can survive crashes. Durable workflow checkpointing lets you resume from an explicit checkpoint ID.
  • Built-in inference is on-device Foundation Models. Inject any InferenceProvider for custom backends; the agent loop stays the same.
  • It is written in Swift all the way down. AsyncThrowingStream, actors, result builders, and macros are first-class here.

Examples

Capability matrix showcase

Swarm now ships with an in-repo capability showcase that exercises the stable surface area in one deterministic matrix:

  • agents and tools
  • streaming
  • conversation plus session persistence
  • sequential, parallel, routed, and repeat-until workflows
  • handoffs
  • memory
  • on-device workspace loading
  • guardrails
  • resilience helpers
  • durable checkpoint and resume
  • observability
  • MCP discovery and tool bridging
  • provider selection

Run it locally:

# Capability showcase matrix covers durable workflows; enable Integrations
swift run --traits Integrations SwarmCapabilityShowcase list
swift run --traits Integrations SwarmCapabilityShowcase matrix
swift run --traits Integrations SwarmCapabilityShowcase run handoff
swift run --traits Integrations SwarmCapabilityShowcase smoke

The deterministic matrix is CI-safe. Live-provider smoke coverage is opt-in through environment variables. See docs/guide/capability-showcase.md for the scenario catalog and smoke-mode details.

End-to-end example apps

Two minimal, buildable apps under Examples/ stress the public API:

| Example | What it proves | | --- | --- | | Examples/OnDeviceChat | Foundation Models chat with @Tool, streaming, and multi-turn Conversation (zero API keys; --demo for CI) | | Examples/MultiAgentPipeline | Sequential + parallel workflows and durable checkpoint/resume (--demo for CI) | | Examples/CodeReviewer | Lightweight CLI that links Swarm and prints a deterministic review plan |

cd Examples/OnDeviceChat && swift run OnDeviceChat --demo
cd Examples/MultiAgentPipeline && swift run MultiAgentPipeline --demo

Optional demos

Package-root demo executables are opt-in so the default library graph stays focused on the framework products:

SWARM_INCLUDE_DEMO=1 swift build
SWARM_INCLUDE_DEMO=1 swift run SwarmDemo
SWARM_INCLUDE_DEMO=1 swift run SwarmMCPServerDemo

Foundation Models First

For Apple platforms, use the built-in on-device path — no API keys:

import Swarm

// Requires macOS/iOS 26+ and Apple Intelligence available on the device.
let agent = try Agent(
    "You are a private on-device assistant.",
    inferenceProvider: .foundationModels()
) {
    // @Tool structs or FunctionTool values
}

let result = try await agent.run("Summarize my notes.")

Notes that matter in production:

  • Availability: use FoundationModelsInferenceProvider.ifAvailable() or check FoundationModelsInferenceProvider.isAvailable before assuming the system model is ready.
  • Tool calling: Swarm bridges @Tool / ToolSchema to Apple's FoundationModels.Tool and executes tools in the agent loop with guardrails intact.
  • Streaming tool calls: not advertised; streaming is text deltas, tools complete as capture-then-execute turns.
  • Dynamic profiles: .foundationModels(profile:) re-resolves instructions/tools/history every turn (WWDC 2026–aligned Swarm API).
  • Linux / CI: Foundation Models is compile-time gated; inject a mock or custom InferenceProvider, or use the deterministic --demo modes in Examples/.

Multi-agent pipeline

let researcher = try Agent("Research the topic and extract key facts.",
    inferenceProvider: .foundationModels()) {
    WebSearchTool()
}

let writer = try Agent("Write a concise summary from the research.",
    inferenceProvider: .foundationModels())

let result = try await Workflow()
    .step(researcher)
    .step(writer)
    .run("Latest advances in on-device ML")

Each agent resolves its own provider. Pass inferenceProvider: per agent (as above), or call await Swarm.configure(provider: myProvider) once at app startup to share a default across every agent that doesn't specify one.

Parallel fan-out

let result = try await Workflow()
    .parallel([bullAgent, bearAgent, analystAgent], merge: .structured)
    .run("Evaluate Apple's Q4 earnings.")
// Three perspectives, merged into one output.

Dynamic routing

let result = try await Workflow()
    .route { input in
        if input.contains("$") { return mathAgent }
        if input.contains("weather") { return weatherAgent }
        return generalAgent
    }
    .run("What is 15% of $240?")

Streaming

for try await event in agent.stream("Summarize the changelog.") {
    switch event {
    case .output(.token(let t)):           print(t, terminator: "")
    case .tool(.completed(let call, _)):   print("\n[tool: \(call.toolName)]")
    case .lifecycle(.completed(let r)):     print("\nDone in \(r.duration)")
    case .lifecycle(.failed(let error)):    print("\nError: \(error)")
    default: break // Other events include .output(.thinking(...)), .handoff(...), .observation(...), and .lifecycle(.iterationStarted(...)).
    }
}

<details> <summary><strong>More examples</strong></summary>

Semantic memory
let agent = try Agent("You remember past conversations.",
    memory: .vector(embeddingProvider: myEmbedder, similarityThreshold: 0.75),
    inferenceProvider: .foundationModels()) {
    // tools
}
Guardrails
let agent = try Agent("You are a helpful assistant.",
    inputGuardrails: [InputGuard.maxLength(5000), InputGuard.notEmpty()],
    outputGuardrails: [OutputGuard.maxLength(2000)])
Closure tools
let reverse = FunctionTool(
    name: "reverse",
    description: "Reverses a string",
    parameters: [ToolParameter(name: "text", description: "Text to reverse", type: .string, isRequired: true)]
) { args in
    let text = try args.require("text", as: String.self)
    return .string(String(text.reversed()))
}

let agent = try Agent("Text utilities.", tools: [reverse])
Crash-resumable workflows
let workflow = Workflow()
    .step(monitor)
    .durable.checkpoint(id: "monitor-v1", policy: .everyStep)
    .durable.checkpointing(.fileSystem(directory: checkpointsURL))

let resumed = try await workflow.durable.execute("watch", resumeFrom: "monitor-v1")
Provider selection
// Built-in: on-device Foundation Models (no API key)
let local = try Agent("Be helpful.", inferenceProvider: .foundationModels())

// Custom backend: any type conforming to InferenceProvider
let custom = try Agent("Be helpful.", inferenceProvider: myCustomProvider)

// Or swap at runtime via environment
let modified = agent.environment(\.inferenceProvider, myCustomProvider)
Conversation
let conversation = Conversation(with: agent)

let response1 = try await conversation.send("What's the weather?")
let response2 = try await conversation.send("And tomorrow?") // Context preserved

for message in await conversation.messages {
    print("\(message.role): \(message.text)")
}

</details>

How Swarm Compares

| | Swarm | LangChain | AutoGen | |---|---|---|---| | Language | Swift 6.2 | Python | Python | | Data race safety | Compile-time | Runtime | Runtime | | On-device LLM | Foundation Models | n/a | n/a | | Execution model | Typed Workflow graph | Loop-based | Loop-based | | Crash recovery | Checkpoints | n/a | Partial | | Type-safe tools | @Tool macro (compile-time) | Decorators (runtime) | Runtime | | Streaming | AsyncThrowingStream | Callbacks | Callbacks | | iOS / macOS native | First-class | n/a | n/a |

What's Included

| | | |---|---| | Agents | Agent struct with @ToolBuilder trailing closure, AgentRuntime protocol | | Workflows | Workflow: .step(), .parallel(), .route(), .repeatUntil(), .timeout() | | Tools | @Tool macro, FunctionTool, @ToolBuilder, parallel execution | | Memory | .conversation(maxMessages:), .vector(embeddingProvider:similarityThreshold:maxResults:), .slidingWindow(maxTokens:), .summary(configuration:summarizer:), .hybrid(configuration:summarizer:) | | Guardrails | InputGuard.maxLength(), InputGuard.notEmpty(), InputGuard.custom(), OutputGuard.maxLength(), OutputGuard.custom() | | Conversation | Conversation actor for stateful multi-turn dialogue | | Resilience | 7 backoff strategies, circuit breaker, fallback chains, rate limiting | | Observability | AgentObserver, Tracer, SwiftLogTracer, per-agent token metrics | | MCP | Model Context Protocol client and server support | | Providers | Built-in Apple Foundation Models (on-device); inject any InferenceProvider for custom backends | | Macros | @Tool, @Parameter, @Traceable, #Prompt |

Architecture

┌─────────────────────────────────────────────────────────────┐
│                      Your Application                       │
│          iOS 26+  ·  macOS 26+  ·  Linux (Ubuntu 22.04+)   │
├─────────────────────────────────────────────────────────────┤
│     Workflow  ·  Conversation  ·  .run()  ·  .stream()      │
├─────────────────────────────────────────────────────────────┤
│  Agents              Memory              Tools              │
│  Agent (struct)      Memory factories    @Tool macro        │
│  AgentRuntime        Conversation        FunctionTool       │
│                      (dot-syntax)        @ToolBuilder       │
├─────────────────────────────────────────────────────────────┤
│  InputGuard · OutputGuard · Resilience · Observability · MCP│
├─────────────────────────────────────────────────────────────┤
│              Durable Graph Runtime (internal)               │
│   Workflow Graph  ·  Checkpointing  ·  Deterministic retry │
├─────────────────────────────────────────────────────────────┤
│              InferenceProvider (pluggable)                   │
│ Foundation Models (built-in) · custom InferenceProvider     │
└─────────────────────────────────────────────────────────────┘

Requirements

| Platform | Minimum | |----------|---------| | Swift | 6.2+ | | iOS | 26.0+ | | macOS | 26.0+ | | tvOS | 26.0+ | | Linux | Ubuntu 22.04+ with Swift 6.2 |

The default Swarm graph is CI-tested on Ubuntu with Swift 6.2. Apple-only features such as Foundation Models, SwiftData, OSLog, and some built-in tool behavior are unavailable or different on Linux; inject a mock or custom InferenceProvider there.

Documentation

| | | |---|---| | Getting Started | Installation, first agent, workflows | | OpenTelemetry Tracing | Export agent and LLM spans, with optional trace header injection for provider HTTP requests | | API Reference | Every type, protocol, and API | | Front-Facing API | Public API surface | | Why Swarm? | Design philosophy and architecture |

Contributing

  1. Fork → branch → swift test → PR
  2. All public types must be Sendable; the compiler enforces it
  3. Format with swiftformat Sources Tests --lint --config .swiftformat

Bug reports and feature requests: GitHub Issues

Community

GitHub Issues · Discussions · @ckarani7

If Swarm saves you time, a star helps others find it.

License

Released under the MIT License.

Package Metadata

Repository: christopherkarani/swarm

Default branch: main

README: README.md