Open source · Swift
AgentKit
A Swift-native SDK for orchestrating agents — multi-model, MCP-integrated, and built so that agent behaviour is testable. It's where my iOS instincts and my curiosity about agents meet: typed, ergonomic, transcript-first.
What it is
AgentKit gives Swift developers a first-class way to build agentic systems — define a session, give it tools, talk to one or more model providers, and get a structured transcript back. It treats the conversation transcript as the central data structure rather than an afterthought.
The design seams
- @SessionSchema macro — declare an agent session's shape and get type-safe structure generated for you.
- SimulatedSession — test agent logic deterministically without hitting a live model — agents you can actually unit-test.
- Multi-model + MCP — swap providers behind one interface; integrate Model Context Protocol tools natively.
- Proxy-based auth — keys never ship in the client; calls route through a proxy.
Why I built it
The fastest way to learn a domain is to build the tool you wish existed in it. AgentKit forced me to reason about agent orchestration, tool calling, evaluation and provider abstraction from first principles — in the language I know best. That's just how my curiosity works: I understand a system by building it.