A fast, keyboard-summoned native macOS quick-chat panel built with SwiftUI and AppKit.
Summon a lightweight, distraction-free conversation window from anywhere with Option + Space, chat with OpenAI (via your existing local Codex CLI login) and open-weight models on OpenRouter from a single unified model picker, and dismiss right back to your work.
Zero external dependencies, zero external Swift packages, and no Electron/WebView or Node.js wrapper.
- Unified Model Picker (OpenAI + OpenRouter)
- OpenAI: Uses your local
codex loginsession (gpt-6-lunadefault when available) with hosted live web search—no OpenAI API key required. - OpenRouter: Optionally add an OpenRouter API key in Settings (
⌘,) to chat with fast, personable open-weight models alongside OpenAI, or run standalone with OpenRouter alone. - Curated Open-Weight Defaults: Ships with DeepSeek V4.1 Flash, Mistral Large 4, Qwen 3.8 Flash, Kimi K3, GLM 5.3 FlashX, Gemma 4 26B A4B, and DeepSeek V4 Pro—customizable directly in Settings (
slugorslug | Display Name). - Seamless Mid-Chat Switching: Switch between OpenAI and OpenRouter models inside the same conversation without losing context.
- Beam Follow-Ups: Preserves Beam's required reasoning metadata privately across turns and app restarts. Older replies without that metadata are supplied as quoted conversation context; reasoning is never shown in chat or forwarded to other models.
- Readable Errors: Failed requests expose a Details popover with the HTTP status, available provider diagnostics, selectable text, and Copy.
- OpenAI: Uses your local
- On-Demand Web Search
- Hosted live web search on OpenAI models and on-demand OpenRouter web search (
plugins: [{ id: "web" }]) triggered only when current information is needed, with inline Markdown source citations.
- Hosted live web search on OpenAI models and on-demand OpenRouter web search (
- Native macOS Experience
- Compact, draggable, resizable
NSPanelwith semantic light/dark appearance (clean white chat surface in light mode, native contrast in dark mode). - Joins all Spaces and works over fullscreen apps; stays open when it loses focus. Dismiss it with Escape or Command-W while keeping your draft, conversation, and warm backend ready.
- Native selectable prose, code blocks with one-click copy, and readable Unicode math normalization.
- The menu-bar bubble includes About SpeedChat, which shows the version, build number, the time the app bundle was built in the local time zone, and the copyright notice from the bundle.
- Compact, draggable, resizable
| Shortcut | Action |
|---|---|
| ⌥ Space (configurable) | Show or focus the SpeedChat panel from anywhere |
| ↵ (Enter) | Send message |
| ⇧↵ (Shift + Enter) | Insert newline in composer |
| ⌘N / ⌃N | Start a new chat (keeps selected model) |
| Esc | Stop active streaming response, or hide the panel when idle |
| ⌘W | Hide the panel without interrupting an active response or clearing draft |
| ⌘, | Open Settings |
| ⌘Q | Quit SpeedChat |
| Page Up / Page Down | Scroll transcript |
- macOS 14 (Sonoma) or newer (Apple Silicon /
arm64). - At least one model provider:
- OpenAI via Codex CLI: Install the Codex CLI and sign in once with
codex loginin Terminal, and/or - OpenRouter: An API key from openrouter.ai, configured in SpeedChat Settings (
⌘,) (works standalone with zero CLI dependencies).
- OpenAI via Codex CLI: Install the Codex CLI and sign in once with
When using OpenAI via Codex, the app automatically discovers codex (and node if your codex shim uses it) across /opt/homebrew/bin, /usr/local/bin, ~/.local/bin, common version-manager paths (~/.volta/bin, ~/.asdf/shims, ~/.local/share/mise/shims, ~/.fnm/current/bin, ~/Library/pnpm), and ~/.hermes/tools without sourcing shell startup files.
For Codex, auto-discovery prefers ~/.local/bin/codex when installed, so an older system-wide copy does not mask the user installation's model catalog. An explicit Codex path in Settings or SPEEDCHAT_CODEX takes precedence.
OpenAI model choices come from the selected Codex CLI's catalog. If newer models are missing, update that Codex installation and restart SpeedChat. If you have multiple installations, check the Codex path in Settings; an explicit path continues to use that copy.
- Download
SpeedChat-macos-arm64.zipfrom the latest GitHub Release, unzip it, and dragSpeedChat.appinto/Applications. - Because the prebuilt app is ad-hoc signed rather than notarized with a paid Apple Developer ID, strip macOS's browser quarantine flag once in Terminal before first launch:
xattr -dr com.apple.quarantine /Applications/SpeedChat.app
open /Applications/SpeedChat.appClone the repository, run the Swift verification suite, build the signed .app bundle locally, and install to /Applications:
scripts/test.sh
scripts/build-app.sh
scripts/smoke-ui.sh
scripts/install-app.shscripts/build-app.shcompiles the release Swift executable, renders the.icnsapp icon natively with AppKit, bundles resources intodist/SpeedChat.app, stampsSpeedChatBuildDatein Info.plist with the UTC build time, and applies an ad-hoc code signature for local personal use. A launch without that key (for example an unpackagedswift run) shows a development build instead of a placeholder date.scripts/install-app.shvalidates the bundle signature and identifier (local.speedchat), stages the update, gracefully quits any running instance, retains a rollback copy (SpeedChat.previous.app), installs/Applications/SpeedChat.app, and launches it without altering your saved chats, preferences, or login items.scripts/package-release.shbuilds and packagesdist/SpeedChat-macos-arm64.zip(with SHA-256 checksum) for attaching to a GitHub Release. Packaging usesscripts/build-app.sh, so the zip includes the same build timestamp, and it stops if that timestamp is missing.
To run directly from the checkout without installing to /Applications:
open dist/SpeedChat.app- Open SpeedChat (⌥Space) and press ⌘, (or click the menu-bar bubble icon → Settings…, or choose Configure OpenRouter in Settings… from the header model menu).
- Paste your OpenRouter API key (
sk-or-v1-...) using ⌘V or the Paste button. - Customize the open-source model list if desired:
- Add any OpenRouter model slug (
provider/model-nameorprovider/model-name | Custom Name). - Remove models you don't use with ×, or click Reset Defaults at any time.
- Add any OpenRouter model slug (
- Pick any model under the OpenAI or OpenRouter section of the header menu.
- Chat & Web Search Scope Only: Local file tools, shell execution, MCP servers, plugins, skills, browser automation, and code execution are explicitly disabled. Codex runs in an isolated empty temporary directory (
0700) with aread-onlysandbox andapprovalPolicy: "never". - Credential Isolation:
- Conversation state and OpenRouter configuration are stored in
~/Library/Application Support/SpeedChat/(state.jsonandopenrouter.json) with directory mode0700and file mode0600. - Subprocess environments are strictly allowlisted—ambient API keys,
NODE_OPTIONS, and OpenRouter credentials are never inherited by the Codex child process.
- Conversation state and OpenRouter configuration are stored in
- Safe Rendering & Links: Assistant replies render as native SwiftUI text/attributed strings (never HTML/WebView). All clicked links are validated at the native boundary (
LinkPolicy), permitting only well-formedhttp/httpsURLs without embedded credentials or control characters. Clipboard actions useNSPasteboarddirectly.
To remove local state after quitting:
rm -rf "$HOME/Library/Application Support/SpeedChat"# Run the full Swift Testing suite (ChatCore + NativeBridge)
scripts/test.sh
# Run the native UI fixture smoke test (generates docs/ui-smoke/report.json)
scripts/smoke-ui.shSee docs/verification.md for detailed test coverage and verification notes.
See THIRD_PARTY_NOTICES.md for third-party software notices.
Licensed under the MIT License. Copyright (c) 2026 Dana Pellerin.