A native macOS app for multi-LLM chat, side-by-side compare, debate, and consensus-finding — all in one window.
One Mac app. Every LLM. Chat with one, compare two, watch them debate, or make them reach consensus.
⬇️ Download the .dmg — prebuilt, no compiling
English | 中文
Stop agonizing over "which model is best." Summon them all in one window — Claude, GPT, Gemini, DeepSeek, GLM, Qwen, Moonshot, your own gateway. Have them answer independently, line them up side by side, make them debate each other, or force them to reach consensus.
Chat · ⌘1 · Pick one model, chat normally |
Compare · ⌘2 · Same prompt, two models streaming side by side |
Debate · ⌘3 · Lead vs Critic, 1–4 rounds, pause anytime |
Discuss ✨ · ⌘4 · Protocol-driven cross-correction, auto-detect convergence |
One place to wire up every model · global + China LLMs + 6 web-search backends |
|
- 🪟 A real .app. Native menu bar, Dock badge, status item, ⌘N/W/Q shortcuts, close-without-quit — full macOS feel.
- 💬 Four conversation modes, one keystroke to switch:
- Chat — talk to a single model, just like ChatGPT
- Compare — same prompt, two models streaming side by side
- Debate — Lead vs Critic, 1–4 rounds, pausable, adoptable mid-flight
- Discuss ✨ new in v0.2 — both sides cross-question via protocol; once both confidence ≥ 8 and verdicts match, it converges early
- 📝 Export Plan ✨ new in v0.3 — one button distills a finished discussion into a plan file written straight into your project folder, reviewed by a second model before it lands, with a ready-to-paste kickoff prompt for Claude Code / Codex CLI
- 🌐 Connect anything. Anthropic API + OpenAI API + Gemini + Claude CLI (subscription) + Codex CLI (subscription) + any OpenAI-compatible endpoint
- 🔍 Built-in web search, 6 backends to pick from (Tavily / Exa / Brave / Serper / Jina / Bocha) — decoupled from the LLMs
- 🔐 API keys live in macOS Keychain; the JSON config only stores templates and metadata
- 📦 Local-first. SQLite + FTS5 full-text search (CJK trigram), every conversation stays on your machine
- 🌍 English & Chinese, 319 strings at 100% coverage
- 🚀 Zero front-end build. DaisyUI v5 + Tailwind v4 + Alpine.js, all CDN — edit, refresh, done.
git clone https://github.com/birdindasky/ask-mac.git
cd ask-mac
make build && make dmg
open dist/Ask-0.5.0.dmgDrag Ask into Applications, launch via Spotlight (Ask) — drop in one model's API key in Settings and you're chatting.
Don't want to build? Grab a prebuilt
.dmgfrom Releases.
Data lives in ~/Library/Application Support/Ask/, logs in ~/Library/Logs/Ask/ask.log,
API keys go straight into the system keychain (service com.birdindasky.ask).
Ask isn't signed with an Apple Developer cert and notarized (that's a $99/year club fee 😅), so the first double-click will be blocked by macOS with a dialog like:
"Ask can't be opened because Apple cannot check it for malicious software."
Don't panic — this does not mean Ask is malware, just that I haven't paid the Apple tax. Three ways around it (pick one, only needed once):
- In Finder, locate
Ask.app(in /Applications or wherever you dragged it) - Control-click (or right-click) Ask.app → choose Open
- Click Open again in the popup
After that, double-click works normally.
If you already double-clicked and got blocked:
- Open System Settings → Privacy & Security
- Scroll to the bottom — you'll see "Ask was blocked because it is from an unidentified developer"
- Click Open Anyway, enter your Mac password
- After this, double-click works fine
xattr -dr com.apple.quarantine /Applications/Ask.appStrips the quarantine attribute — double-click works forever after.
⚠️ This applies to every unsigned / non-Mac-App-Store Mac app, not just Ask.
git clone https://github.com/birdindasky/ask-mac.git
cd ask-mac
python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
python run.py
# open http://127.0.0.1:8870 in your browserIn dev mode data lands in ~/.ask-dev/ and keys fall back to JSON instead of the keychain (handy for resets).
| Mode | Shortcut | What it does |
|---|---|---|
| Chat | ⌘1 | Pick one model, normal multi-turn chat |
| Compare | ⌘2 | Pick two models, same prompt streams to both in parallel |
| Debate | ⌘3 | Lead + Critic (or symmetric), 1–4 rounds, pausable, adoptable |
| Discuss ✨ | ⌘4 | Protocol-style 6-field exchange, sides correct each other, auto-detects consensus |
Open ⚙️ Settings → "Web Search", pick a backend, paste the key — done.
| Backend | Sign up | Notes |
|---|---|---|
| Tavily | tavily.com | Built for RAG, returns summaries, free 1000/month |
| Exa | exa.ai | Neural search, great at digging up deep content |
| Brave Search | api.search.brave.com | Independent index, free 2000/month |
| Serper.dev | serper.dev | Google results, super fast, free 2500 |
| Jina Search | jina.ai/reader | Returns markdown, LLM-friendly |
| Bocha | bochaai.com | China-based AI search, optimized for Chinese |
One backend is enough. Any of them works with every LLM provider — the flow is: prompt → search backend → top-N hits → glued into the system context → fed to the LLM. The model itself doesn't need to understand any search-tool protocol.
If you're already logged into a Claude Code or OpenAI Codex subscription in your terminal, you can run them as CLIs directly:
- In Settings, pick the Claude CLI (subscription) or OpenAI Codex CLI (subscription) template — no API key needed
- Ask spawns
claude/codexas a subprocess with all sneaky billing-leak env vars already scrubbed (ANTHROPIC_API_KEY,OPENAI_API_KEY,*_BASE_URL, etc.)
Prerequisite: you can already run
claudeorcodexin your terminal and you're signed into the subscription.
In Settings, pick "OneAPI / NewAPI self-hosted" or "Custom OpenAI-compatible", then fill in:
base_url: e.g.http://127.0.0.1:3000/v1api_key: the key your gateway issued- Model list: one per line, the model IDs your gateway has enabled
Environment variables (active in source mode; the .app uses defaults):
| Variable | Default | Notes |
|---|---|---|
MLC_HOST |
127.0.0.1 |
uvicorn bind address |
MLC_PORT |
8870 |
port |
MLC_DATA_DIR |
dev ~/.ask-dev/ .app ~/Library/Application Support/Ask/ |
config + database |
MLC_LOG_DIR |
dev ~/.ask-dev/logs/ .app ~/Library/Logs/Ask/ |
log directory |
MLC_CLI_TIMEOUT |
120 |
per-request CLI timeout (seconds) |
MLC_PACKAGED |
0 |
1 forces .app paths (testing) |
MLC_FORCE_KEYCHAIN |
0 |
1 makes dev mode also use the keychain |
MLC_RELOAD |
0 |
1 enables uvicorn hot reload |
source venv/bin/activate
python -m pytest -qCovers: adapter registration, env scrub, config migrations, session/message CRUD, and full SSE flows for all four modes. 42 cases.
mac_launcher.py .app entry: uvicorn-in-thread + PyWebView + AppKit chrome
run.py dev-mode entry
setup.py py2app config
Makefile dev / test / icon / build / dmg / install
scripts/
├── build_icon.py generates assets/Ask.icns
└── build_dmg.py packages dist/Ask.app into a .dmg
assets/
├── Ask.icns app icon
└── Ask.iconset/ multi-resolution icon source (auto-generated)
app/
├── main.py FastAPI wiring / static-file path resolution
├── settings.py paths and constants (dev vs .app switch)
├── db.py SQLite + FTS5 messages_fts
├── config_store.py config.json persistence (API keys never land in JSON)
├── api/ REST + SSE routes
├── modes/ chat / compare / debate / discuss
├── providers/ per-vendor adapters + cli_detect PATH resolution
├── security/ keychain wrapper + secrets helpers
├── search/ 6 web-search backends + citation injection
└── utils/ token_budget / attachments / autostart / notifier / dock_badge
static/ front end (single-page HTML + Alpine + DaisyUI + Tailwind)
tests/ 42 pytest cases
- One-click self-update — Settings → Version & updates checks GitHub for a newer release, shows the changelog, and updates the app with one button. It downloads the DMG itself (no Gatekeeper quarantine), swaps the bundle in place with a backup-and-restore safety net, and relaunches — no manual drag. A dismissible banner announces new versions on launch. You install this version once by hand; every release after updates in-app.
- Liquid Glass theme — the dark theme becomes Apple's frosted-glass look: a near-black ground lit by colored halos, translucent blurred panels everywhere, gradient buttons, clean tail-less bubbles. All four modes and every dialog inherit it; the light theme stays clean for daytime use.
- Vendor brand avatars — every answer now carries the real logo of the model behind it (Anthropic / OpenAI / Gemini / DeepSeek / GLM / Qwen / Kimi / Doubao / MiniMax / Yi / Groq and more). Identified by model-id, so aggregators still show the true brand. Logos are bundled locally.
Export Plan (出计划) — Ask used to be a meeting room where conclusions died on the whiteboard: you copied text out by hand and pasted it into a coding agent. Now there's a button.
- Tag a session with a project (📁 chip in the toolbar) — exports go straight to that folder; untagged sessions ask at export time.
- Writer + reviewer, two models — pick any two wired models; the writer distills the whole discussion into a 7-section plan (
background / settled conclusions / open disagreements / steps / risks / acceptance / source), the reviewer checks the draft against the transcript — dropped conclusions, open questions passed off as settled, invented content — and can send it back for up to 2 rewrites. Both turns stream into the session as labeled bubbles. - Approved → written to
<project>/docs/ASK-PLAN-YYYYMMDD-<topic>.md(never overwrites; the file carries a "do not commit" banner) plus a one-tap kickoff prompt with the absolute path — paste it into Claude Code or Codex CLI and the work starts. - Rejected → nothing is written. The objections land in the session, and in Discuss mode "Continue" feeds them into the next rounds so the models answer the reviewer directly. The fix loop is more discussion, not a defective file.
- Plan files can only land inside your home directory, one file per export, no git operations ever.
- Discuss mode — a fourth tab where two models alternate in a protocol-style 6-field format (
[Current verdict] / [Confidence] / [Support] / [Corrected by other] / [Still holding] / [Need from other]). The backend reads[Confidence]numbers in real time — both ≥ 8 + verdicts agree → converge early; otherwise run the full N rounds (default 3, adjustable 1–5). When it ends, side A appends a summary with a📌 Consensusbadge and a📋 Copy consensusbutton. - Context progress bar — token usage estimated after each turn; near 90% you get a "Compact history" button.
- Regenerate — hit ↻ on the last assistant bubble to regenerate, or pick a different model from the dropdown to redo the answer with that model.
- ⌘F global search — SQLite FTS5 search across every conversation; trigram tokenizer means Chinese hits work too.
- Attachments — drop images or text files into the chat box, works in all four modes.
- Welcome wizard — 4-screen first-launch tour that gets your first key wired up in 30 seconds.
- i18n at 100% — English & Chinese, 319 strings, switch in Settings.
- Full desktop chrome — NSMenu / NSStatusItem tray / Dock badge / system notifications / launch-at-login / About panel.
- Close-without-quit — closing the main window just hides it; ⌘Tab brings it back; Dock Quit /
killall/ Activity Monitor all clean-shutdown properly.
If you fork this for your own use, strongly consider changing birdindasky / Ask in the three places below to your own identifiers — otherwise two apps will fight over the same macOS Keychain entry and data directory:
APP_NAMEandBUNDLE_IDinapp/settings.pyCFBundleName/CFBundleDisplayName/CFBundleIdentifierin the plist insidesetup.pyAPP_NAME(and the resultingDMG_NAME) in theMakefile
Then make build && make dmg produces your own standalone .app, with data in ~/Library/Application Support/<your APP_NAME>/, fully isolated from the original author's environment.
- Image attachments are currently base64-buffered and the LLM only sees an
[attached image: name]placeholder; real vision support lands in v0.3 via Anthropic / OpenAI / Gemini multimodal APIs. - No code signing / notarization (built for personal use, no Apple Notary). For real distribution, configure a Developer ID and edit
setup.py. - v0.2 has no built-in auto-update — upgrades mean re-running
make dmg.
The hands-on acceptance checklist lives in ACCEPTANCE.md.
MIT © 2026 — use it, modify it, ship it commercially, redistribute it. Just don't sue me.
Built by birdindasky · Designed for vibecoders who hate switching tabs
If this project helped you, drop a star 🌟 to make the author's day.