StreamingText makes LLM responses in your SwiftUI app feel like ChatGPT or Claude: tokens flow in at a steady, human rhythm, half-written Markdown never flickers, and the text never falls behind the network.

@State private var reply = StreamingTextModel()

StreamingTextView(reply)
    .task { try? await reply.stream(client.streamResponse(for: prompt)) }

Screenshots

Captured from the example chat app on an iOS 26 simulator by CI.

Mid-stream (healed Markdown + caret)	Finished answer	Dark mode

Why

Token streams are bursty. A naive text += delta makes the UI stutter: nothing for 400 ms, then a whole sentence at once, with raw ** and ` characters blinking in and out while Markdown is incomplete. StreamingText sits between your network layer and your view and fixes all of that.

Features

• ⌨️ Adaptive pacing: a steady typing speed that automatically speeds up when the backlog grows, so the display lags at most catchUpDuration seconds.
• 🩹 Markdown healing: closes open **bold, *italic, ~~strike and `code on every frame, and hides markers that have no content yet.
• 🧬 Grapheme safe: emoji, flags and combining accents are never split, even across chunk boundaries.
• 🔌 Works with any AsyncSequence: OpenAI, Anthropic, Apple Foundation Models, your own SSE parser.
• ⏹ Skip & stop: revealAll() finishes instantly; cancelling the task stops the stream.
• ▍ Blinking caret (.block, .dot, .bar or your own) while text is active.
• ♿️ VoiceOver reads the full text, not the half-typed frame.
• 🧵 Swift 6 strict concurrency, @Observable, zero dependencies, tested with Swift Testing.

Installation

Add the package in Xcode via File › Add Package Dependencies…:

https://github.com/halilozel1903/StreamingText

or in Package.swift:

.package(url: "https://github.com/halilozel1903/StreamingText", from: "1.0.0")

Usage

Stream an AsyncSequence

import StreamingText

struct AnswerView: View {
    let prompt: String
    @State private var answer = StreamingTextModel(pacing: .natural)

    var body: some View {
        ScrollView {
            StreamingTextView(answer, caret: .dot)
                .frame(maxWidth: .infinity, alignment: .leading)
                .padding()
        }
        .task {
            try? await answer.stream(MyLLMClient.stream(prompt))   // any AsyncSequence<String>
        }
    }
}

Push chunks yourself

answer.beginReceiving()
for try await event in sseEvents {
    answer.append(event.delta)
}
answer.finishReceiving()

Pacing presets

Preset	Speed	Max lag
.natural	45 chars/s	1.5 s
.fast	120 chars/s	0.6 s
.instant	no animation	0 s
.init(charactersPerSecond:catchUpDuration:)	your call	your call

Messages from history

StreamingTextView(StreamingTextModel(text: savedMessage.body))

Model state

Property	Meaning
displayedText	What is on screen right now
fullText	Everything received so far
isReceiving	The source is still producing text
isRevealing	Buffered text is still being typed
isActive	Either of the above; the caret shows while this is true

Use the pieces on their own

Both building blocks are public, pure value types:

MarkdownHealer.heal("This is **impor")        // "This is **impor**"
MarkdownHealer.heal("Hello **")               // "Hello "

var buffer = StreamingTextBuffer(pacing: .fast)
buffer.append("Hello")
buffer.advance(by: 1 / 60)                    // reveal what is due this frame

Example app

Example/ contains a small chat app with a fake LLM that streams bursty tokens, a speed picker and a stop button. Generate the project with XcodeGen:

brew install xcodegen
cd Example && xcodegen generate
open StreamingTextDemo.xcodeproj

Requirements

• Xcode 26+ (Swift 6.2)
• iOS 17+ / macOS 14+

License

MIT. See LICENSE.