How do I install H/Ai Browser?

Copy the stdio config for Cursor, Claude Desktop, Claude Code, or VS Code on this page. It uses the publisher's install, not a PlugMCP proxy.

Does H/Ai Browser need an API key?

No required environment variable is listed on this record.

What transport does H/Ai Browser use?

H/Ai Browser runs locally over stdio.

When was H/Ai Browser last checked?

Last checked 2026-11-09. Let AI agents drive VS Code's built-in browser alongside you, with click-to-source.

H/Ai Browser MCP server

H/Ai — hai-browser

Share VS Code's integrated browser with Claude Code, GitHub Copilot, Codex CLI, Gemini CLI, Cline or any other MCP agent. You and the agent work in the same browser tab: it reads the page, clicks and types with real input, takes screenshots and watches the console, while you watch and take over at any time.

Point at an element and H/Ai opens the line of code that rendered it, and the agent gets the same element (source location, HTML, styles, screenshot), so "make this button match the header" just works.

Status: prototype (phases 0–2). See docs/plan.md.

How it works

Claude Code ──stdio MCP──▶ hai-browser-mcp ──ws://127.0.0.1 (token)──▶ H/Ai VS Code extension
                                                                          │ editor-browser debug session
                                                                          │ + js-debug requestCDPProxy
                                                                          ▼
                                                               VS Code integrated browser tab (CDP)

The extension attaches to an integrated-browser tab through VS Code's built-in JavaScript debugger (editor-browser debug type) and asks it for a Chrome DevTools Protocol proxy (extension.js-debug.requestCDPProxy). This uses only stable, public APIs, so it works in published extensions.
Each VS Code window runs a localhost-only WebSocket server and writes a lockfile (~/.hai-browser/sessions/<pid>.json, mode 0600) with its port and a random token.
hai-browser-mcp picks the window whose workspace contains the agent's working directory and forwards tool calls to it.

Everything runs on your machine. There is no hosted service and no telemetry.

Requirements

VS Code 1.119+ (integrated browser + built-in JavaScript debugger). VSCodium and Windsurf work too; they install H/Ai from Open VSX. Cursor and Antigravity don't yet: Cursor ships without VS Code's integrated browser, and Antigravity is based on VS Code 1.107.
Claude Code, GitHub Copilot, Codex CLI, Gemini CLI, Cline or another MCP client
Node.js 20+ to build from source (installed users can do without it; H/Ai falls back to VS Code's runtime)

Install

Install the extension: search for H/Ai Browser in VS Code's Extensions view, or run code --install-extension hai-browser.hai-browser. You can also download hai-browser.vsix from the latest release and run code --install-extension hai-browser.vsix, or use the one-line installer:
```
# macOS / Linux
curl -fsSL https://raw.githubusercontent.com/SnapBlock/hai-browser/main/scripts/install.sh | sh
```
```
# Windows (PowerShell)
irm https://raw.githubusercontent.com/SnapBlock/hai-browser/main/scripts/install.ps1 | iex
```
Open a folder in VS Code and trust it (the extension does not run in Restricted Mode). H/Ai asks "let Claude Code use this browser?". Click Connect Claude Code. You can run it again any time with H/Ai: Connect Claude Code.
Start claude in that folder (VS Code's terminal is easiest) and ask it to open your app, e.g. "open http://localhost:5173 and check the signup form works". Or share a tab you already have open with H/Ai: Share Browser Tab with Agent (also in the status bar).

The MCP server ships inside the extension. Connecting copies it to ~/.hai-browser/mcp/index.mjs (kept up to date when the extension updates) and runs claude mcp add --scope user hai-browser -- node ~/.hai-browser/mcp/index.mjs. If Node.js 20+ isn't on your PATH, it uses VS Code's own runtime instead.

Other agents

GitHub Copilot (agent mode in VS Code): nothing to set up. H/Ai registers its MCP server with VS Code, so its tools show up in the chat's tools list.
Codex CLI, Gemini CLI, Cline and other MCP clients: run H/Ai: Connect Other Agents, pick your agent, and H/Ai copies the setup to your clipboard. It's the same server either way (the path exists once H/Ai has run in VS Code):
```
codex mcp add hai-browser -- node ~/.hai-browser/mcp/index.mjs
gemini mcp add --scope user hai-browser node ~/.hai-browser/mcp/index.mjs
```
For JSON-configured clients such as Cline, add {"mcpServers": {"hai-browser": {"command": "node", "args": ["/Users/<you>/.hai-browser/mcp/index.mjs"]}}} with your real home path.

Setup from source

pnpm install
pnpm build
pnpm --filter hai-browser package   # packages/extension/hai-browser.vsix

Install that VSIX (or press F5, "Run hai-browser extension", in this repo) and connect as above. Pushing a v* tag that matches the extension's version builds the VSIX, attaches it to a GitHub release, and publishes it to the VS Code Marketplace (repo secret VSCE_PAT) and Open VSX (OVSX_PAT) when those secrets are set.

The npm plugins publish separately through npm trusted publishing (no token): bump the plugin's version, then push a tag named <package>@<version>, e.g. [email protected]. .github/workflows/npm-publish.yml checks the tag matches and runs npm publish.

Jump from the page to the code

Add the dev-only plugin for your framework (React/JSX; production builds are untouched).

Vite

npm install -D hai-browser-vite

// vite.config.ts
import react from '@vitejs/plugin-react';
import hai from 'hai-browser-vite';

export default defineConfig({ plugins: [hai(), react()] });

Next.js (Turbopack or --webpack, App or Pages Router, server and client components)

npm install -D hai-browser-next

// next.config.ts
import { withHaiBrowser } from 'hai-browser-next';

export default withHaiBrowser({ /* your config */ });

withHaiBrowser only changes the config under next dev. It adds a turbopack.rules loader for *.jsx/*.tsx (skipping node_modules) and, for webpack dev builds, the same loader as an enforce: 'pre' rule. Other webpack setups can use the loader directly: { test: /\.[jt]sx$/, exclude: /node_modules/, enforce: 'pre', use: ['hai-browser-next/loader'] }.

It tags each HTML element with where it was written, e.g. <button data-hai-src="src/PlanCard.tsx:6:7">. Components are not tagged themselves (they may not pass unknown props to the DOM), so you land on the JSX element inside the component.

Then run H/Ai: Pick Element and Open Source (status bar Pick while a tab is shared), hover to highlight, and click. The click is captured rather than sent to the page, VS Code opens the file at that line beside the browser, and the element becomes the agent's current selection. Esc cancels.

Agents read it with browser_get_selection, or call it with wait: true to ask you to pick something. Snapshots also show src=file:line:col next to each tagged element.

Replay an agent run

H/Ai saves each agent run on your machine, step by step: the tool the agent called and its inputs, the element it acted on (with its source file when the Vite/Next plugin is on), where it clicked, a screenshot after the step, and the result or error.

Run H/Ai: Replay Agent Run (or click Replay in the status bar) and pick a run. Step through it with ← →, press Space to play, or drag the slider. Click a step's source to open it in the editor. Export HTML saves the run as one file that opens in any browser, so you can attach it to a PR or send it to a reviewer.

One run per agent session; a new run starts after 10 minutes without activity. Runs are deleted automatically after 7 days, and only the newest 20 are kept, up to 200 MB in total.
Runs stay on this machine (VS Code's extension storage). Nothing is uploaded.
Text typed into password, API key, code and card fields is not saved, and that text is hidden wherever it shows up later in the run. Anything shaped like an API key or token (sk-…, live_…, ghp_…, JWTs and similar) is hidden in the saved steps too. Screenshots are saved as they were, so a key shown on the page is still visible in them.
Turn recording off with the haiBrowser.recordRuns setting, and delete saved runs with H/Ai: Delete Recorded Agent Runs.

Tools

Tool	What it does
`browser_status`	Whether a tab is shared, plus URL and title
`browser_open`	Open a URL in the shared tab. With none shared, reuses an open tab already showing that URL, else opens a new one
`browser_request_share`	Ask the user to pick one of their open tabs to share
`browser_navigate`	Go to a URL, or back / forward / reload
`browser_snapshot`	Accessibility-style outline with `[ref=eN]` handles
`browser_click`	Click by ref or `x`/`y` (trusted mouse input); right/middle button, double click, modifier keys
`browser_hover`	Move the mouse over a ref or point
`browser_scroll`	Mouse-wheel scroll at a ref, point or the viewport centre, or scroll a ref into view
`browser_drag`	Press, move and release between two refs/points (sliders, sortable lists)
`browser_type`	Type into an element by ref, optionally clear and submit
`browser_press_key`	Press a key or chord (`Enter`, `ArrowDown`, `Control+A`, `Shift+Tab`, …)
`browser_select_option`	Pick `<select>` options by value or label
`browser_upload_file`	Set a file input's files (paths must be inside the workspace)
`browser_handle_dialog`	Accept or dismiss an `alert` / `confirm` / `prompt`
`browser_wait_for`	Wait for text to appear or disappear, or for some seconds. Text that appears is highlighted for you
`browser_show`	Scroll to text or an element, point the cursor at it and highlight it with an optional caption, so you see what the agent read (codes, links, prices)

Every action (click, hover, scroll, drag, type, press_key, select_option, upload_file, handle_dialog, wait_for) returns the page snapshot afterwards, like a computer-use agent seeing the screen after each step. Pass screenshot: true to also get an image, or snapshot: false to skip the outline. While the agent works, its pointer glides to each spot it clicks, hovers, drags or types into, with a ripple on each click, so you can follow along. A small purple face rides with it: its eyes look where the pointer is heading, blink now and then, and squint on each click. It never receives events, is left out of snapshots and is hidden from the agent's screenshots; turn it off with haiBrowser.showAgentCursor. If an action opens a JavaScript dialog, the result says so and other tools refuse to run until browser_handle_dialog handles it.

Tabs are the ones shared with agents (opened by browser_open/browser_tab_new, or shared by you). VS Code's browser-tab API is still proposed, so agents can't see or switch to your other tabs until you share them.

FAQ

Is it free? Yes. H/Ai is MIT-licensed and free. You bring your own agent (a Claude Code subscription, an API key, and so on); H/Ai adds no cost and needs no account.

Does anything leave my machine? Not through H/Ai: there is no hosted browser, no server of ours, and no telemetry. Your agent sends what it reads (snapshots, screenshots, logs) to its own model provider, as with any other tool it uses.

Which agents work? Any MCP client. Claude Code, Codex CLI and Gemini CLI have been tested end to end on a real store checkout. GitHub Copilot agent mode picks up the server automatically. Cline and others use the JSON config above. The server is also on npm (npx -y hai-browser-mcp) and in the MCP Registry.

Which editors work? VS Code 1.119+, VSCodium and Windsurf. Cursor doesn't yet, because it ships without VS Code's integrated browser. Antigravity doesn't either; it's based on VS Code 1.107.

Gemini CLI won't sign in with my Google account. Google no longer allows Gemini CLI's Google sign-in for individual accounts. Create a free key at AI Studio and set GEMINI_API_KEY.

Which frameworks support click-to-source? React with Vite (hai-browser-vite) and Next.js with Turbopack or webpack (hai-browser-next). On other pages, picking still gives the agent the element's HTML, styles and screenshot, just not the source line. Vue and Svelte are next; open an issue if you need one.

How is this different from Playwright MCP or Chrome DevTools MCP? Those usually launch and drive a separate browser. H/Ai uses the tab inside VS Code that you're already looking at: you see the agent's cursor, can take over at any time, and can point at an element to show the agent what you mean.

Can the agent see my other tabs? No. Only tabs the agent opened or that you shared with H/Ai: Share Browser Tab with Agent.

Security

The agent API listens on 127.0.0.1 only and rejects connections without the per-window token.
Snapshots and picked-element HTML never include password values or autocomplete="cc-*" fields.
Arbitrary JavaScript (browser_evaluate) is disabled unless you turn it on.
Page content is untrusted: a web page can contain text written to manipulate an agent (prompt injection). Share tabs you trust, and prefer a separate profile for anything sensitive.
Before a risky action on a website, H/Ai asks you: typing into a password, API key, code or card field; submitting a form; typing (or putting in an address) text the agent read on a different site; or acting on a page with text that looks like orders for an AI. Choose Allow once, Allow on this site (no more questions there until the agent disconnects) or Stop. Reading and browsing never ask, and neither do pages on localhost. Turn it off with haiBrowser.confirmActions.
When an agent signs up for things or reads codes from email, give it a throwaway inbox (the demos use Poof), not your own.
The status bar shows H/Ai: sharing whenever a tab is shared; click it to stop.

Development

pnpm typecheck && pnpm build && pnpm test
python3 -m http.server 8765 -d examples/demo   # demo page
pnpm smoke                                     # drives the demo through the MCP server (needs VS Code + extension running)

pnpm --filter example-vite-react dev           # React demo tagged by hai-browser-vite, on :5173
node packages/mcp/scripts/computer-smoke.mjs http://127.0.0.1:8765/ /abs/path/in/workspace.txt   # hover, drag, dialogs, uploads, tabs…
pnpm --filter hai-browser-mcp smoke:pick       # agent asks for a pick, checks it maps to src/PlanCard.tsx

pnpm --filter example-next-app dev             # Next.js demo (Turbopack) on :3000; or dev:webpack (webpack, :3001), one at a time
pnpm --filter example-vite-shop dev            # multi-page store (search, filters, cart, checkout, forms) on :5180
HAI_PICK_APP=next pnpm --filter hai-browser-mcp smoke:pick   # same check against app/Counter.tsx

Packages:

packages/extension — VS Code extension (browser bridge + agent API)
packages/mcp — hai-browser-mcp stdio MCP server
packages/vite-plugin — hai-browser-vite, dev-only data-hai-src tagging
packages/next-plugin — hai-browser-next, the same tagging for Next.js (Turbopack and webpack)
packages/protocol — shared types for the extension ↔ MCP protocol
examples/demo, examples/vite-react, examples/next-app — test pages
examples/vite-shop — a realistic multi-page store for end-to-end agent runs

License

MIT

H/Ai Browser

Add this server

Host a copy with our keys and logs.

Server facts

Tools

Questions

How do I install H/Ai Browser?

Does H/Ai Browser need an API key?

What transport does H/Ai Browser use?

When was H/Ai Browser last checked?

Documentation

H/Ai — hai-browser

How it works

Requirements

Install

Other agents

Setup from source

Jump from the page to the code

Replay an agent run

Tools

FAQ

Security

Development

License