ZenCopy 发布常驻桌面 AI 智能体，双击 Ctrl/Cmd+C 唤起对话

Instant AI. Right on your screen.

Talk to an AI anywhere, the moment you want to — the answer to whatever you copied appears on the spot, and the conversation continues right there. A calm, resident desktop agent, summoned by Ctrl/Cmd + C + C.

What it is

ZenCopy is a resident agent that summons an AI conversation on top of any app. No chat site to open, no pasting — you ask about what you are looking at, right where you are.

The signal is Ctrl/Cmd + C + C — your ordinary copy, twice, quickly. From there, in order:

what you had selected — text, image, or files — is captured,
a small popup opens in the corner of your screen,
the default prompt streams its answer immediately (out of the box that is Summarize, which distills the copy to its essence in one sentence; switching to Explain, Translate, Polish, or your own prompts takes one click or one number key),
and a message field waits under the answer — "more detail", "as a table", "in English?". The conversation continues in place, no window switch.

A single normal copy is never disturbed.

Use it for a while and a habit forms: when you want to ask an AI, you just copy twice. Opening a chat site, finding the tab, pasting — all of it disappears. The copy everyone already knows becomes the door to an AI.

Why ZenCopy

Plenty of tools put an AI popup near your selection — most are macOS-only, text-only, and tied to one vendor's models or a subscription. ZenCopy's combination is what's rare:

The conversation comes to you — no hunting for a chat window; the first answer is automatic, the rest is yours to steer.
Windows, macOS, and Linux, one behavior.
Your model, your rules — any provider (OpenAI / Google / Anthropic / a corporate gateway / local Ollama); keys never leave your machine, and a free Gemini API key is enough to start.
Everything you copy — text, images, and files, not just selected text. You can even talk to a screenshot.
Rules you control — which prompt runs is decided by declarative rules (capture kind, plus app / window / URL / length overrides), edited in the GUI or as a file.
Calm and honest — no account, no subscription, no telemetry. Apache-2.0.

Highlights

Conversation in place — reply under any answer with questions, instructions, or tangents; every reply carries copy and retry (retry rewinds the thread to that reply), Enter sends, Shift+Enter breaks a line, and an IME's confirming Enter never sends by accident.
One global trigger — Ctrl/Cmd + C + C, anywhere. Powered by the copycopy crate.
Streaming, calm popup — the copied content on top, the operation in the middle, the result below.
Provider-agnostic LLM — built on the Vercel AI SDK + ai-sdk-catalog. Roles map to models in a single config file; swap providers without touching prompts.
Not just text — formatted copies arrive as Markdown; images and files attach as-is, with the type decided by contents, never by file extension.
Resident & unobtrusive — tray icon, frameless always-on-top popup you can drag and resize, launch-at-login (toggleable), recall the last result from the tray, dismiss with Esc or the close button.
Multi-display & multi-Space aware — windows open where you are working.
Privacy-aware — clipboard content marked sensitive (passwords/secrets) is ignored.
Theme & i18n — light / dark / system; 19 languages including RTL (Arabic, Persian, Hebrew), following the OS by default.

Platforms: Windows, macOS, and Linux. macOS needs 13.3 (Ventura) or later. On Linux, GNOME on Wayland is supported via a bundled GNOME Shell extension (installed automatically, no sudo) and X11 via a key listener; other Wayland compositors (KDE, wlroots, …) are not supported yet. The Linux packages are built on Ubuntu 22.04 and need glibc 2.35 or later.

The browser's address — the URL condition of a rule, {{ url }} in a prompt — is read on Windows and macOS. macOS asks for it once per browser, the first time you copy there (Automation, in Privacy & Security); declined, everything else works and the address stays empty. Linux offers no way to read it.

Develop

Prerequisites: Bun, the Rust toolchain, and the Tauri v2 system dependencies.

bun install
bun run tauri dev      # run the app
bun run tauri build    # produce a release bundle

Frontend checks: bun run lint, bun run build, bun run format. Backend: cargo build / cargo clippy / cargo fmt in src-tauri/.

Configuration

Settings split into two layers:

GUI (Settings window) — AI provider (a free Gemini key is the fastest start), the prompts editor, rules (per-kind assignments and override rules), popup position, theme, language, launch at login. Stored per-user.
Files (power users) — read from the per-user app config dir (macOS: ~/Library/Application Support/app.zencopy/; the exact path is logged at startup). Defaults for rules and prompts are embedded in the app, so these files only ever override:
- ai-sdk-catalog.json — providers and role → model mapping. Only default is required. Each pre-installed prompt also answers to a role named by its id (zencopy-summarize, zencopy-explain, …): map one and that prompt runs on its own model, leave it out and the prompt runs as default. API keys live inline in this local file (what the settings UI writes) and never leave your machine.
- rules.json — kind → prompt, plus higher-priority overrides.
- prompts/*.md — prompt definitions: YAML frontmatter (id, label, optional role, instructions) + a Liquid prompt body. Template variables like {{ text }} embed the capture's context — the full table lives in the configuration docs.

The Settings window covers the common setups (OpenAI / Google / Anthropic / any OpenAI-compatible endpoint such as Ollama); editing ai-sdk-catalog.json unlocks the rest — multiple providers, corporate gateways, per-role models.

Privacy

No telemetry, no accounts, no intermediary server — copied content goes only to the LLM provider you configure. See PRIVACY.md.