One browser shared by a human and an AI — the human sees the rendered page, the AI reads the DOM and page source. Same tabs, same live session, at the same time.
English | 中文
Duplex start page (dark theme). A light theme is also included.
The AI drives: it searches Bilibili, opens the video and starts playback — the side panel streams every step.
Full demo — 3 min, bilingual subtitles, with music:
One continuous walkthrough: Act 1 — the built-in agent searches Bilibili and plays a paper-explainer video; Act 2 — Kimi browses to Claude and asks what CUDA is, then summarizes the answer; feature highlights — multi-model APIs, skills, external agents, emergency stop; and where to get it.
Controlling Microsoft Excel for the web — no dedicated spreadsheet-editing skill installed; only the stock API and a few simple automation skills 😅
A feasibility demo, not a recommended workflow. With no task-specific skill to lean on, Duplex worked with the stock API plus a few simple automation skills alone — improvising clipboard read/write channels, DOM probing and the like to push a whole grade sheet into Excel for the web. It got there, but the process wasted a lot of time and tokens. 😅
-
Security boundaries (high stakes — please read). Duplex lets the AI drive your real browser session, including login state, cookies and local data. Under this architecture, a wrong AI move can touch real accounts and data (sending messages, submitting forms, modifying or deleting content), potentially with serious consequences. Don't leave the agent running unattended in environments where sensitive accounts are signed in.
Esctakeover is a last-resort human brake — it is no substitute for your own judgement about what the AI should be allowed to touch. -
The tooling layer is still being tuned. The goal is for the human to stay in command of the AI's tools, rather than have AI tooling and the human's own actions crowd each other out of the pipeline (contending for the same page, interrupting input, interleaving conflicting actions). The trade-offs here are still evolving — feedback on human/agent contention is welcome via Issues.
- Skills — Claude Code-compatible skill format (
SKILL.md): indexes your existing~/.claude/skillsand supports importing or authoring duplex-only skills under~/.cobrowse/skills. The built-in agent reads and follows them on demand (every script run asks for confirmation first). - Multi-protocol model APIs — beyond OpenAI-compatible, native support for Anthropic Messages, OpenAI Responses and Google Gemini.
- Credential import — reuse your local Codex (ChatGPT subscription) and opencode logins directly; no API key typing needed.
- External agent tools — switch the panel between opencode / Codex / Claude Code / Gemini CLI / Qwen Code / custom CLIs: transcripts mirror into the panel (history + live), Codex / Claude Code sessions can be continued by replying right from the panel (the message headlessly resumes the same session), and new sessions can be started right from the panel. Codex / Claude Code also get a model picker (default follows the CLI config; one click writes the choice back to the global config, with a timestamped backup); custom tools can be added manually (session directory + start command).
- Safety & robustness — global emergency-stop hotkeys (default
Esc/F2, customizable from the ⌨ button), triple-confirm deletes, task watchdog with automatic render recovery, process-tree termination, and more.
| Duplex | browser-use | Browser MCP | Playwright MCP | AI browsers (Atlas / Comet) | |
|---|---|---|---|---|---|
| Form | Desktop browser (Electron app) | Python automation framework | MCP server (browser extension) | MCP server (Microsoft) | Closed-source product |
| Who uses the browser | Human and AI share the same tab and the same live session | AI only (separate automation instance) | AI drives your current Chrome | AI only (Playwright instance) | AI assistant alongside/operating |
| What the AI sees | DOM outline snapshot + source + screenshots | Vision + DOM | Screenshots + a11y tree | Accessibility tree | Internal |
| Human collaboration | Real-time side-by-side; Esc takeover interrupts the AI anytime |
Logs afterwards | Human spectates | Human spectates | Limited intervention |
| Connectable AI | Built-in models + opencode / Codex / Claude Code / Gemini / Qwen / any MCP client | Bring your own LLM | Any MCP client | Any MCP client | Official model only |
| Conversation visibility | Live side-panel mirror (including external CLIs' chats and tool calls) | Logs/terminal | In the client | In the client | In-app |
| Data | Fully local | Local/cloud | Local | Local | Cloud |
In one line: browser-use / Playwright MCP let the AI run a flow for you; Duplex lets the AI use the browser together with you — same tab, same session, takeover anytime.
- A real browser — tabs, address bar with search (Baidu / Bing / Google), back / forward / reload, loading state, themes (light / dark / follow system) and a wallpaper start page with a clock and search.
- 24 MCP tools for AI agents —
snapshotcompresses any page into a compact DOM outline with[eN]refs; the other tools cover tabs, navigation, clicking, typing, dragging, file upload, scrolling, waiting, console logs, JS evaluation and page annotations. - Zero-setup bridge — a stdio MCP bridge (
mcp-bridge) auto-launches the browser on the first tool call. Works with opencode, Claude Code, or any MCP client. - Live session mirror — when your AI works through opencode, its replies, reasoning and tool-call cards stream into the side panel in real time. Type in the panel to inject a message into the same session.
- AI action visualization — a translucent cursor, element highlight and a status bar ("AI is clicking «…» — Esc to take over") are drawn in a Shadow-DOM overlay, so you always see what the AI is doing on the page.
- Esc takeover — press
Esc(or click the status bar) to take control instantly: the running tool call is aborted, in-flight waits return early, and the AI is told the user took over. - Page annotations — press
?or useannotation_modeto draw a box / circle / arrow / point on any page and attach a question. The annotation is compiled into a structured text brief (DOM outline + visible text + selectors + geometry) and sent to the AI. - Built-in agent (optional) — connect any OpenAI-compatible API (DeepSeek, Kimi, Qwen, GLM, Ollama, …) and let the browser drive itself. Provider management supports one-click import from opencode.
- Conversation history — built-in agent sessions are saved locally and can be reopened from the history menu.
The start page shows a clock, date and search box over a wallpaper that follows the active theme (light and dark wallpapers switch automatically with the system).
Ask the built-in model in the side panel — in Chinese or English. It calls tools (navigate, snapshot, …), with each call shown as a card, and reports back in the panel while you watch the page change.
When your AI runs through opencode, its messages and tool calls are mirrored into the panel while it drives the same visible browser. The "opencode" / "built-in" tabs at the top of the panel switch between the two ways of working.
Pick which opencode session the panel is connected to. "Auto" follows the most recent conversation.
Manage OpenAI-compatible providers for the built-in agent: add, edit, delete, or import providers directly from your opencode configuration. Local models via Ollama (http://localhost:11434/v1) work out of the box.
Built-in agent conversations are stored locally and can be reopened at any time.
Every AI action is drawn on the page: a cursor ring, an element highlight and a status bar. Press Esc at any moment to take the browser back — the AI stops immediately and waits for your instruction.
Draw a box (or circle / arrow / point) around anything and ask a question about it. The annotation is converted into a structured text brief for the AI, including the DOM outline of the region, visible text, selectors and geometry — so even text-only AIs can "see" what you mean.
- Download the latest installer from Releases:
Duplex Setup x.y.z.exe(Windows),.dmg(macOS, Apple Silicon / Intel) or.AppImage(Linux). - Run it and launch Duplex.
npm install
npm run build # Electron app -> out/
npm run build:bridge # stdio MCP bridge -> dist-bridge/index.cjsIf Electron's binary download fails (e.g. behind a slow mirror), retry with:
$env:ELECTRON_MIRROR="https://npmmirror.com/mirrors/electron/"; node node_modules/electron/install.js
Run it:
npm run dev # development mode with HMR
# or
npm start # preview the production build- Install the mirror plugin. Copy
integrations/opencode/plugins/cobrowse-mirror.tsto your global opencode plugin directory (~/.config/opencode/plugins/), or into.opencode/plugins/of a project. - Register the MCP server in your opencode config (
opencode.json), project-level or global — seeopencode.example.json:
{
"mcp": {
"duplex": {
"type": "local",
"command": ["node", "C:\\path\\to\\Duplex\\dist-bridge\\index.cjs"],
"enabled": true
}
}
}- In opencode, ask the AI to browse: "open example.com and describe the page". The browser launches automatically on the first tool call (no need to start it manually).
- In the Duplex side panel choose the opencode tab and select the session you want to follow (or keep "Auto").
Panel → Built-in tab → Model settings → add an OpenAI-compatible provider (base URL, API key, model name), or click Import from opencode to reuse your existing provider configuration. For local models, point the base URL to Ollama, e.g. http://localhost:11434/v1.
| Tool | Description |
|---|---|
list_tabs / new_tab / close_tab / switch_tab |
Tab management — human and AI share the same tabs |
navigate / search / history |
Open a URL (auto-searches if not a URL), search (Baidu default, or Bing / Google), back / forward / reload |
snapshot |
Page as text — compact DOM outline; interactive elements carry [eN] refs |
get_html / query |
Raw HTML, or detailed info for elements matching a CSS selector |
screenshot |
PNG screenshot of the visible page (for multimodal models) |
click / dblclick / hover |
Click / double-click / hover; target accepts an eN ref or a CSS selector |
type / press |
Type text (optionally submitting) and press keys, incl. combos like Control+A |
drag |
Real drag & drop from one point/element to another |
select_option / upload |
Native <select> options, and file upload by absolute path |
scroll |
Scroll the page or a specific element into view |
wait |
Wait for time / selector / text — interruptible by Esc |
get_console |
Read page console logs (errors and warnings) |
annotation_mode |
Enter/exit the annotation overlay (human draws a box/circle/arrow/point + question) |
evaluate |
Evaluate JS in the page, returns JSON-serializable results |
- The Electron main process serves a small local HTTP API on
127.0.0.1protected by a per-launch bearer token (endpoint info is written to~/.cobrowse/endpoint.json). dist-bridge/index.cjsis a stdio MCP server that proxies tool calls to that API and auto-launches the app when it is not running.- The opencode plugin pushes session events (text, reasoning, tool calls) into the panel, and long-polls for messages queued in the browser (~10 ms injection latency).
- Panel messages and page annotations are injected into the active opencode session with
session.promptAsync. - Codex and Claude Code are mirrored from their local session transcripts (history + live tail); replies sent from the panel headlessly resume the same session (
codex exec resume/claude --resume) and the new turns stream back through the same tail. - External scripts can also push messages into the session:
POST /api/chat { "text": "..." }.
- Page internals inside iframes and shadow DOM are not covered by
snapshot/click— shadow DOM is only detected, not entered. eNrefs are invalidated by navigation; re-runsnapshotafter the page changes.- Message injection targets the "active session"; with several opencode sessions the target may occasionally be ambiguous.
- The mirror store holds the recent event stream in memory; it resets on browser restart.
- The built-in agent is a convenience option: local / smaller models are noticeably less reliable at long tool-use chains than a full opencode setup.
- Richer client integrations — Codex / Claude Code support has landed; more on the way:
- The bridge is client-agnostic (standard stdio MCP), so the tool layer is not opencode-specific by design.
- Session mirroring is in: transcripts replay into the panel live, and replies sent from the panel resume the same session headlessly (Codex
exec resume/ Claude Code--resume) — the CLI appends to its transcript and the tail streams it back into the panel. - Next: richer event mapping (tool calls / reasoning cards) and more clients (Gemini CLI / Qwen Code).
npm run dev # dev mode (electron-vite, HMR for the renderer)
npm run typecheck # TypeScript checks (node + web)
npm test # unit tests (vitest)
npm run smoke # end-to-end smoke test (launches the bridge and a real browser)
npm run dist # build the Windows installer (electron-builder)Debug helpers:
GET /api/debug/ui-snapshot— captures the current window to~/.cobrowse/ui-snapshot.png.POST /api/debug/panel-eval— runs JS in the panel renderer (only when the app is started withCOBROWSE_DEBUG_UI=1).- Logs:
~/.cobrowse/app.log(main + renderer log lines).