webview-cli

Native macOS UIs for CLI AI agents.
_{~290KB · ~180ms cold start · no Electron · no npm · no runtime}

curl -sSL https://raw.githubusercontent.com/giannimassi/webview-cli/main/install.sh | bash

One command. Installs the binary via Homebrew, installs the Claude Code skill if ~/.claude exists, runs a smoke test.

---

~290KB vs Electron's 50MB+

Your AI agent needs to ask you a question that doesn't fit in a terminal prompt. The choices today:

• Electron app: 50MB+ runtime, 500–800ms cold start, requires a persistent app bundle. Way too heavy for a subprocess the agent spawns 20 times per session.
• Tauri / Wails: 8–15MB, 300–500ms. Still too heavy, still designed as app frameworks.
• osascript AppleScript dialog: built-in, ~300ms, gives you 2 buttons and a text field. Useless for anything structured.
• Terminal [y/N]: the default. Fine for yes/no, terrible for "review this diff" or "pick one of these 12 PRs."

webview-cli: ~290KB single binary, ~180ms cold start, a real native macOS window, structured JSON back to the agent, process dies clean.

	webview-cli	Electron	Tauri / Wails	osascript
Binary size	~290KB	50MB+	8–15MB	built-in
Cold start	~180ms	500–800ms	300–500ms	~300ms
Rich HTML/CSS UI	✅	✅	✅	❌
Structured JSON out	✅	app-specific	app-specific	❌ (string)
Spawnable subprocess	✅	❌ (app lifecycle)	❌ (app lifecycle)	✅
Runtime deps	none	Electron bundle	WebView2/WebKit2 + Rust	none
Agent-first design	✅	❌	❌	❌

Why it's this small

The numbers aren't optimization magic. They come from not bundling things that were already there.

WKWebView is already on your Mac. The same engine Safari uses. We wrap it instead of shipping Chromium. That's ~50MB saved, no install-time download, and a decade of Apple's work on rendering correctness and memory management comes for free.

The Swift runtime ships with macOS 12+. libswiftCore, libswiftFoundation, and friends live on the OS. We don't bundle them, version-pin them, or worry about ABI skew. Another ~12MB avoided, and none of the dylib-shipping pain that usually comes with cross-platform Swift.

A2UI is declarative, not imperative. Your agent describes what UI it wants in a flat JSONL stream; the renderer decides how to lay it out, handle focus, and collect form data. ~250 lines of vanilla JavaScript covers the entire catalog — no React, no templating engine, no build step.

Each subtraction compounded. Together they produced a binary that starts in ~180ms — faster than Electron finishes parsing its own bootstrap scripts.

Who it's for

Agent developers on macOS.

Claude Code was macOS-first. Cursor is macOS-first. Codex CLI, ChatGPT Desktop, Raycast, Warp, the wave of indie MCP servers being published on GitHub every week — the people building and using agent workflows skew heavily toward macOS. If you've opened Claude Code recently, you're probably holding a Mac.

macOS-only is a choice, not a gap. A Linux port (GTK + WebKit2) and a Windows port (WebView2) are both technically straightforward, and the protocol is platform-independent — but shipping them at v0.1 would mean three build pipelines, three sets of edge cases, and slower iteration on the platform the audience actually uses. Depth over breadth. If you need this on another OS badly enough to maintain a port, open an issue — a clean implementation gets merged.

---

Use it

With Claude Code

Install the bundled skill once:

ln -s "$(pwd)/skill" ~/.claude/skills/webview   # or use the installer flag

Then ask your agent anything that needs structured input from you. The skill handles the rest — generates A2UI, spawns the binary, parses the result, returns typed data. You never write JSONL by hand.

Example: deploy approval. Agent says:

"I'm about to deploy payment-service to production. Should I proceed?"

The skill opens a native window with the change summary, a rollout radio (canary / full), a comment field, and Approve/Cancel. You click. The agent receives:

{"action": "approve", "data": {"rollout": "canary", "note": "Monitoring on standby"}}

and continues. No terminal input. No context loss.

Other patterns the skill handles out of the box: single-select from agent-found options (PRs to review, branches to rebase), multi-field config forms, diff-and-acknowledge flows. Full skill docs: skill/SKILL.md.

With OpenAI Codex CLI, Gemini CLI, or any subprocess-capable agent

The binary is protocol-agnostic. Anything that can spawn_subprocess and read stdout works:

import subprocess, json

result = subprocess.run(
    ["webview-cli", "--a2ui", "--timeout", "120"],
    input=your_a2ui_jsonl,
    capture_output=True, text=True
)
# result.returncode: 0=submitted, 1=cancelled, 2=timeout, 3=error
# json.loads(result.stdout) has {"status", "data": {"action", "data": {...}}}

A complete wrapper plus a Codex tool definition you can paste into your agent config: examples/openai-codex-tool.md.

From a shell script or MCP server

It's a Unix tool. Stdin in, stdout out, exit codes report outcome. Pipes work. Blocking works. Put it wherever a subprocess can run:

# Approval gate in a CI/deploy script
RESULT=$(cat my-form.jsonl | webview-cli --a2ui --title "Deploy?" --timeout 300)
case $? in
  0) ACTION=$(echo "$RESULT" | jq -r '.data.action'); [ "$ACTION" = "approve" ] && deploy ;;
  1) echo "User cancelled." ;;
  2) echo "Timed out — no response in 5 min." ;;
esac

---

Common patterns

Pick from options Agent enumerates candidates (PRs, branches, files). User picks one with radio buttons. Agent proceeds with the choice.	Multi-field config Text inputs + selects + checkboxes in one native form. Better than six read -p prompts in a row.
Destructive confirmation Irreversible action with context, safety checkbox, and a danger-variant button. Prevents the "yes 40 times in a row" mistake.	Triage / classification Priority dropdown, complexity radio, and notes in one native form. Agent gets structured metadata back to act on.
Markdown review Agent generates a spec or doc. User reviews with paragraph-level comments, optionally edits the source, and submits structured feedback. Agent reads the review and can iterate. Full example	Combine markdown review with form Review a spec AND pick options in one window. Comments, edits, and form fields together. Full example

What the raw A2UI JSONL looks like (click to expand — the skill writes this for you)

{"surfaceUpdate":{"components":[{"id":"root","component":{"Column":{"children":{"explicitList":["card"]}}}}]}}
{"surfaceUpdate":{"components":[{"id":"card","component":{"Card":{"children":{"explicitList":["title","diff","risk","note","btns"]}}}}]}}
{"surfaceUpdate":{"components":[{"id":"title","component":{"Text":{"usageHint":"h2","text":{"literalString":"Deploy payment-service to prod?"}}}}]}}
{"surfaceUpdate":{"components":[{"id":"diff","component":{"Text":{"usageHint":"body","text":{"literalString":"3 files · +47/-12 · CI green · ENG-1234"}}}}]}}
{"surfaceUpdate":{"components":[{"id":"risk","component":{"RadioGroup":{"label":{"literalString":"Rollout"},"fieldName":"rollout","options":[{"value":"canary","label":"Canary (10%)"},{"value":"full","label":"Full rollout"}]}}}]}}
{"surfaceUpdate":{"components":[{"id":"note","component":{"TextInput":{"label":{"literalString":"Deploy note"},"fieldName":"note","multiline":true}}}]}}
{"surfaceUpdate":{"components":[{"id":"btns","component":{"Row":{"alignment":"end","children":{"explicitList":["c","go"]}}}}]}}
{"surfaceUpdate":{"components":[{"id":"c","component":{"Button":{"label":{"literalString":"Cancel"},"variant":"secondary","action":{"name":"cancel"}}}}]}}
{"surfaceUpdate":{"components":[{"id":"go","component":{"Button":{"label":{"literalString":"Deploy"},"variant":"success","action":{"name":"approve"}}}}]}}
{"beginRendering":{"root":"root"}}

That's the whole approval UI above. One line per component, flat adjacency list, LLM-friendly to generate. The skill produces this from a short natural-language description of the UI.

---

Install

One command (recommended)

curl -sSL https://raw.githubusercontent.com/giannimassi/webview-cli/main/install.sh | bash

Handles Homebrew tap + formula install + Claude Code skill install + smoke test. Flags:

• --with-claude-skill — force skill install even if ~/.claude not detected
• --no-claude-skill — skip skill install

Homebrew only

brew tap giannimassi/tap
brew install webview-cli

From source

git clone https://github.com/giannimassi/webview-cli.git
cd webview-cli
make install   # copies to ~/bin/webview-cli

Requires macOS 12+ and the Swift toolchain (Xcode Command Line Tools is enough).

---

Supported A2UI components

Text, TextInput, Button, Column, Row, Card, Select, Checkbox, RadioGroup, Image, Divider. Subset of Google's A2UI v0.8 standard catalog.

Full prop reference: docs/a2ui-subset.md. Protocol reference: docs/protocol.md. Architecture tour: docs/architecture.md.

---

How it works

One Swift file, ~550 lines. NSApplication with .accessory policy (no Dock icon), one NSWindow with a WKWebView, WKScriptMessageHandler bridging JS events to stdout, WKURLSchemeHandler serving an embedded renderer via agent://. Stdin feeds A2UI JSONL into the renderer.

The renderer itself is ~250 lines of vanilla ES — no React, no framework, no build step. It's embedded as a string literal in the Swift binary. "What CSS framework is that?" is a frequent question. The answer is none — system fonts, -apple-system, CSS custom properties, ~60 lines of hand-written styles.

See docs/architecture.md for the tour.

---

License

MIT — see LICENSE.

Credits

• A2UI by Google — the declarative UI spec this renders a subset of
• Opus 4.7 — this tool was built in one evening with Claude Code driving the keyboard. It pairs well with its builder.

---

Built by @giannimassi. If this unblocks one of your agent workflows, a star is appreciated.