TextKit

TextKit is a native macOS menu bar utility for clipboard-first text work. Copy text anywhere on macOS, open TextKit from the menu bar, then rewrite, summarize, extract, reply, prompt, or reduce the copied text locally.

The app is built for a frictionless non-technical install path: packaged builds bundle the local llama.cpp runtime, prompt users through first-run model setup, and run inference on-device after the selected model has been downloaded.

Current Product

• Menu bar app with an icon-only status item and SwiftUI popover.
• Six top-level tools: Rewrite, Prompt, Extract, Reply, Summarize, and Reduce.
• Four modes per tool, including Clean, Short, Professional, Bullet, Brief, Executive, Logs, Structured, and related task-specific modes.
• Automatic clipboard intake with safeguards for app-authored clipboard copies.
• Manual Reduce flow for long text, logs, traces, and structured blobs.
• Local model setup UI with progress, model status, and retry support.
• Advanced prompt profile editor with preview, reset, import/export, strict mode, temperature, max tokens, and seed controls.
• Settings controls for warm runtime behavior, downloaded model cleanup, app data reset, and uninstall.
• Sparkle-backed update checks for signed release builds.

Local AI

TextKit currently supports two local model options:

• Stable: Qwen/Qwen2.5-0.5B-Instruct-GGUF
• Experimental: AaryanK/Qwen3.5-0.8B-GGUF

The app installs one balanced GGUF file per selected model for the user-facing setup flow. The Response Mode setting changes generation behavior, not the downloaded model file:

• Fast: shorter results and lower effective token budget.
• Balanced: default for everyday clipboard work.
• Quality: allows more detail and a larger effective token budget.

When warm cache is enabled, TextKit prefers a local llama-server worker and shuts it down after the configured idle window. If the warm worker is unavailable or disabled, TextKit falls back to one-shot llama-completion.

Reduce is intentionally local-rule based and does not require the model.

Privacy Model

TextKit has no cloud inference path. Copied text is processed locally by the app and the bundled local runtime. Network access is only needed to download a selected model during setup.

User Install Flow

For prerelease or release builds:

1. Open the DMG or ZIP-provided app.
2. Move TextKit.app into Applications.
3. Open TextKit from Applications.
4. On first run, choose a model in the setup flow.
5. Download the model once, then use TextKit offline for normal text operations.

Packaged builds do not require Homebrew on the user's Mac.

Development Requirements

TextKit is a Swift Package macOS app.

• macOS 26 target
• Xcode 26 toolchain
• Swift 6.2+ package manifest
• Full Xcode install expected at /Volumes/SSD/Applications/Xcode.app/Contents/Developer for the local scripts in this workspace

Build and launch the local app:

./script/build_and_run.sh

Run tests:

xcrun swift test

The local build script stages the app at:

dist/TextKit.app

Developer Model Setup

The developer setup script uses Homebrew llama.cpp to populate a local model cache for development:

./script/setup_model_runtime.sh

Install the experimental model for local testing:

./script/setup_model_runtime.sh --model experimental --quant balanced

Run a model smoke test:

./script/setup_model_runtime.sh --smoke-test

End-user release builds should use the bundled runtime path instead of requiring Homebrew.

Golden Eval Harness

TextKit includes a golden eval harness for rewrite behavior with tuned development cases and separate holdout cases.

Run the default dev suite:

./script/run_golden_eval.sh

Run the holdout suite:

./script/run_golden_eval.sh --suite holdout

Run a single mode or case:

./script/run_golden_eval.sh --mode rewrite.short
./script/run_golden_eval.sh --case board-update

Compare the experimental model against the same suite:

./script/run_golden_eval.sh --model experimental
./script/run_golden_eval.sh --model experimental --suite holdout

Measure the rewrite shaping layer by ablating rewrite heuristics:

./script/run_golden_eval.sh --suite holdout --ablation rewriteHeuristics --threshold 0.0

Packaging

Build a local prerelease package:

./script/package_release.sh --version 0.1.0

The release script:

• builds the release app
• bundles the local llama.cpp runtime
• packages ZIP and DMG artifacts
• embeds Sparkle for in-app update checks
• generates a signed Sparkle appcast when SPARKLE_PRIVATE_ED_KEY is provided
• supports optional Developer ID signing and notarization when Apple credentials are provided through environment variables

Generated artifacts are written under:

dist/release/

See docs/distribution.md for GitHub Actions, signing/notarization secrets, release artifacts, and Homebrew tap publishing.

Repository Layout

• Sources/App: app entrypoint and menu bar delegate
• Sources/Views: popover, setup, and settings UI
• Sources/Models: tools, modes, model options, and request types
• Sources/Stores: app state, settings, and output cache
• Sources/Services: clipboard monitor, routing, prompt composition, inference, setup, reduction, and eval support
• Sources/Support: runtime lookup, cleanup, process runner, icons, and window helpers
• Tests/TextKitTests: unit tests, golden eval tests, and fixtures
• script: build, setup, eval, packaging, and runtime bundling scripts
• Resources: app icon resources

License

TextKit is distributed under the AGPL-3.0 license. See LICENSE.