Site: https://annotkit.gpu-cli.sh

A Swift package for native in-app annotation, built for AI coding agents. Click a UI element in your own macOS or iOS app, or drag a frame around it, attach a note, and emit an agent-readable, code-locating annotation. The native analogue of the web Agentation tool.

The gesture becomes a stable selector, an element path, the element's role and text, and your comment, so an AI coding agent can locate the exact view instead of guessing from a verbal description.

Requirements: Swift 6.1 toolchain (Xcode 16.3 or later), macOS 15+ or iOS 17+. The package is Swift 6 language mode with strict concurrency. It has no dependencies.

Xcode: File ▸ Add Package Dependencies…, paste https://github.com/gpu-cli/annotkit, choose the version rule Up to Next Major from 0.8.0, and add the AnnotKit library to your app target. Leave AnnotKitMCP unchecked unless you want the agent bridge; the toolbar does not need it.

Package.swift:

dependencies: [
    .package(url: "https://github.com/gpu-cli/annotkit", from: "0.8.0")
],
targets: [
    .executableTarget(
        name: "MyApp",
        dependencies: [.product(name: "AnnotKit", package: "annotkit")]
    )
]

AppKit (macOS): in applicationDidFinishLaunching, after your main window is up:

import AnnotKit

func applicationDidFinishLaunching(_ notification: Notification) {
    // ... make and show your window ...
    #if DEBUG
    Annotation.install()   // floating toolbar; click a view, type a note
    #endif
}

If several windows are open at launch, name the one to annotate instead of letting AnnotKit pick: Annotation.install(on: window).

SwiftUI (macOS): call it from the root view's .onAppear or from an NSApplicationDelegateAdaptor:

@main
struct MyApp: App {
    var body: some Scene {
        WindowGroup {
            ContentView()
                .onAppear {
                    #if DEBUG
                    Annotation.install()
                    #endif
                }
        }
    }
}

SwiftUI (iOS): attach it to the root view:

ContentView()
#if DEBUG
    .installAnnotation()
#endif

UIKit iOS hosts call Annotation.install() from the scene delegate once the window is key.

install() is a no-op in release builds unless the ANNOTKIT_ENABLE environment variable is set, and a no-op in debug builds when ANNOTKIT_DISABLE is set, so the #if DEBUG above is belt and braces: the toolbar never appears in a normal shipping build either way.

By default notes are written to ANNOTKIT_NOTES.md in the process's current working directory. For an app launched from Xcode that is usually not your project folder. Either set the working directory in the scheme (Edit Scheme ▸ Run ▸ Options ▸ Working Directory), point the path at the repo through the scheme's environment variables (ANNOTKIT_NOTES_MD=/path/to/repo/ANNOTKIT_NOTES.md), or pass a sink with an explicit path:

Annotation.install(sink: NotesFileSink(path: "/path/to/repo/ANNOTKIT_NOTES.md"))

Other sinks: ClipboardSink(format: .markdown | .json) copies instead of writing, JSONFileSink writes the JSON store the MCP bridge reads, and MultiSink fans out to several.

Annotation.install(sink: ClipboardSink(format: .json))

Selectors anchor to the nearest accessibilityIdentifier (#Settings.Models >> @Save). Views with no identifier anywhere above them still get a resolvable selector, but it is built from roles and indices and is fragile across layout changes. Put .accessibilityIdentifier("Settings.Models") on the components you expect to annotate and the notes will point an agent straight at that code.

Register a world-context provider and every captured note snapshots it, so an agent can put back the world the note was made in instead of guessing:

Annotation.install(
    context: { ["persona": currentPersona, "appearance": appearanceName] }
)

Press Annotate on the floating pill, click a control (or switch to the frame tool and drag around a card), type a note, and save. Export writes every pending note through the sink. Run swift run AnnotKitDemo in this repo to try the whole loop in a sample app first.

• macOS queries the app's own accessibility tree (the only strategy that surfaces SwiftUI accessibilityIdentifier values); iOS walks the UIView hierarchy. Both resolve a tap with the same rule: the deepest actionable control wins (a click inside a button binds to the button, not its label glyph), else the deepest meaningful element (a standalone text/label/value leaf is annotated in its own right). The generated selector then anchors a non-identified target to its nearest seeded accessibilityIdentifier (#Settings.Models >> @Save), so it round-trips a resolver and points an agent at the right component's code. See DECISIONS.md.
• Drawing a frame instead of clicking inverts that rule on purpose: a click means "this exact spot" and descends, while a box drawn around a card means "I mean this whole thing", so the largest element the frame surrounds wins and the labels inside it do not. A frame drawn inside something binds to the tightest element enclosing it, and the drawn rect rides along on the note. Saves hunting for the one pixel that hit-tests to a composite component.
• Notes are written in the ANNOTKIT_NOTES.md format that the process-agentation-notes skill consumes, or copied to the clipboard. The file is rewritten with the full set of notes on every export, so it never holds duplicates or stale entries. A screenshot of the element is captured in memory on request (AnnotationSession.screenshotSelected()) but is not part of the exported note; the markdown and JSON carry only pixel dimensions.
• Platform coverage is uneven. The selector engine, target rules, session and sinks are shared and unit-tested on both platforms, and CI cross-compiles for the iOS simulator. The live macOS path is additionally exercised by on-device probes and an end-to-end test; the iOS adapter's live behaviour is covered by unit tests over the pure rules only. See PARITY.md for the per-capability matrix.

Write notes as JSON (JSONFileSink) and run the MCP server so an agent can query pending annotations over the Model Context Protocol:

swift run annotkit-mcp path/to/ANNOTKIT_NOTES.json

Tools: annotation_get_pending, annotation_resolve.

A host that launches many isolated instances of one binary — a design loop with an HMR session per branch, a gallery, an inspect window per persona — configures each of them from the launch environment alone, with no change to the app's call site:

ANNOTKIT_NOTES_MD=$WORLD/notes.md \
ANNOTKIT_NOTES=$WORLD/notes.json \
ANNOTKIT_EVENTS=$WORLD/events.jsonl \
ANNOTKIT_CONTEXT_PERSONA=ada \
  ./MyApp

ANNOTKIT_EVENTS adds an append-only JSONL log beside the snapshot — one line per capture, edit and delete — because the snapshot itself is written atomically and so cannot be followed with tail -f. An agent watches the stream and reads the snapshot the line names:

tail -F "$WORLD/events.jsonl" | while read -r line; do … done

Full contract, including how launcher context and the in-app provider merge: docs/embedding.md.

swift build
swift test
swift run AnnotKitProbe     # live macOS smoke test
swift run AnnotKitDemo      # interactive demo app (overlay mounted; good for recordings)
swift run AnnotKitEnvProbe  # one env-configured embedding host, driven in code

Swift 6 (strict concurrency), macOS 15+, iOS 17+. Cross-compile for the iOS simulator with the command in CONTRIBUTING.md.

MIT. See LICENSE. The note format is a clean-room reimplementation of Agentation's AGENTATION_NOTES.md, written to ANNOTKIT_NOTES.md; this project contains no Agentation source.