Arc is being sunset. Meridian keeps the Arc workflow you're used to (spaces, pinned tabs, a Today list that tidies itself, split view, Peek, the command bar) and imports your Arc sidebar in one click. It's also built around AI from the start: a built-in Claude assistant, and an MCP server so Claude Code, Claude Desktop or any MCP client can read and drive your browser, with access you control.
Features
The Arc workflow
- Spaces with their own colour, icon and profile (separate cookies and logins: work vs. personal). Switch with
Ctrl+1…9 on macOS, Alt+1…9 on Windows/Linux, or ⌘⌥←/→.
- Pinned tabs and folders that remember their home URL. Navigate away and a
/ button takes you back.
- Favorites grid at the top, shared across spaces of the same profile.
- Today: unpinned tabs auto-archive after 12 h (configurable). Everything archived stays searchable.
- Command bar (
⌘T / ⌘L): search open tabs across spaces, history, the archive, commands, or the web.
- Split view (
Ctrl+Shift+=) with a draggable divider.
- Peek: links from pinned tabs to other sites open in an overlay. Expand to a tab or a split.
- Little window: links from other apps open in a small window. "Open in Meridian" moves them in.
- Air Traffic Control (Settings → Links): route links from other apps by domain or URL pattern to a chosen space, or to the Little window.
- Private windows (
⇧⌘N): an in-memory profile; no history, cookies or site data survive closing the window.
- Archive view (
⌥⌘A) to browse, search and restore archived tabs; a downloads popover; a space theme editor with gradients and icons.
- Hide the sidebar (
⌘S) and reveal it by hovering the window edge. Sidebar can sit on the right.
- Drag and drop tabs between favorites, pinned, folders and Today.
- Import from Arc: spaces, colours, icons, pinned tabs, folders, favorites and Today tabs, read locally from Arc's sidebar file on macOS and Windows. On macOS your logins come over too: Arc's cookies are copied into the matching Meridian profiles (after you allow access to Arc's key in the macOS keychain prompt), so sites you were signed in to stay signed in. Nothing leaves your computer and Arc isn't changed.
AI-centric
- Bring your own AI, honestly: Ask, Chat and the welcome tour use the AI you connect: your Claude API key (a question in ⌘T costs a fraction of a cent), GitHub Copilot, or a free local model in Ollama / LM Studio. Settings → AI shows what's installed and signed in and sets it up in one click. With Other agents turned on you can also use Codex, Cursor, OpenCode, Qwen Code, Amp, Goose or Droid with your own account, under each provider's terms. Claude Code and Gemini CLI run in the Workbench terminal: their terms don't allow apps to power features with a subscription login, so Meridian doesn't. Agents answer as text only: no tools, no MCP, no user plugins, the prompt on stdin, in a private empty folder.
- Ask from the command bar (
⌘T): type a question and the answer streams right under it. It knows Meridian itself ("how do I split the screen?"). Continue in chat carries the conversation to ⌘J.
- Modes: Everyday, AI for work (designers, managers, writers) and Developer. The welcome tour asks; the developer tools stay out of sight until you want them.
- Ask panel (
⌘J): summarize, ask about the current page. With a Claude API key it can also act in the browser (open tabs, read pages, click, type).
- MCP server with 21 browser tools:
list_spaces, list_tabs, search_tabs, search_history, read_page, get_interactive_elements, screenshot, open_tab, navigate, history_nav, activate_tab, close_tab, click, type_text, press_key, scroll, wait_for, organize_tab, split_view, create_space, and opt-in run_javascript.
- Tabs an agent is working in glow. An activity bar shows what it did and has a Stop button.
- Mission Control (
⇧⌘M): a live tree of every agent working with Meridian: Claude Code and Codex sessions started from the Workbench (locally or in a Docker sandbox) with their subagents and nested subagents, MCP clients driving the browser, and ⌘J chats. Status (thinking / running a tool / waiting for you / done), the current tool with a redacted summary, elapsed time, tool calls, project. Click to jump to the agent's terminal or tab. A notification and a sidebar badge when an agent waits for you; "N agents working" in the sidebar and the Workbench.
Workbench: your coding agent in a browser tab, with your site right next to it (⇧⌘E, or the code icon at the bottom of the sidebar)
- A tab with a file tree, a Monaco editor (the VS Code editor) and real terminals. One-click launchers for Claude Code, Codex, Gemini CLI or a shell in your project. Put it in split view next to your
localhost preview.
- Agents started here get Meridian's browser tools for that session only, so they can open your app, read it and take screenshots. Your global CLI configs aren't touched.
- Make it yours with agents: the Workbench loads
user.css, settings.json and plugins from its own folder and reloads them live. Click Customize with AI and ask for a bigger font, a deploy button or a new panel. New plugins run only after you approve them. Every change is snapshotted, and Safe Mode starts without customizations.
- Secrets vault: API keys encrypted with your system keychain, global or per project. Agents only see them as environment variables of the processes Meridian starts, never as text. Meridian warns you when you paste a key into a terminal or into ⌘J, offers to move
.env files into the vault, and asks before a Claude Code session reads a secret file.
- Sandboxed agents with Docker or OrbStack: run Claude Code, Codex or a shell in a container that sees only your project. The agent runs as a non-root user with no capabilities.
.git, .claude and other auto-run config are read-only, so the agent can't plant hooks. It gets only the secrets you pick for the session (passed in a file, not visible in docker inspect), and you can switch the network off.
- Hubs: a folder for notes or posts that you write with agents. Create one from the Workbench in one click: it gets a simple structure, notes for agents (
AGENTS.md/CLAUDE.md), git and a starter skill, and is pinned in your sidebar.
Also
- Built-in ad and tracker blocking (Ghostery engine), per-site permission prompts, downloads, find in page, zoom, reopen closed tab, session restore.
- Chromium engine: sites work like in Chrome. Experimental unpacked Chrome extension loading.
Install
Easiest: undermove.github.io/meridian-browser picks the right download for your computer. On macOS and Linux you can also install or update with one line:
curl -fsSL https://undermove.github.io/meridian-browser/install.sh | bash
When a new version is out, New Meridian version available shows at the bottom of the sidebar, like in Arc. On macOS a click installs it and reopens Meridian with your tabs; on Windows and Linux it downloads the new installer.
Or grab a build from Releases:
| Download |
|---|
| macOS | Meridian-…-mac-arm64.dmg (Apple silicon) or …-mac-x64.dmg (Intel) |
| Windows | Meridian-…-win….exe (one installer for x64 and ARM) |
| Linux | Meridian-…-linux-….AppImage, or the .deb for Debian/Ubuntu (x64 and arm64) |
Meridian isn't signed with paid Apple/Microsoft certificates yet, so the first launch needs one extra step:
- macOS: open the
.dmg, drag Meridian to Applications, then run once in Terminal:
xattr -dr com.apple.quarantine /Applications/Meridian.app
(or try to open it, then System Settings → Privacy & Security → Open Anyway).
- Windows: if SmartScreen says "Windows protected your PC", click More info → Run anyway.
- Linux:
chmod +x Meridian-*.AppImage && ./Meridian-*.AppImage, or sudo apt install ./Meridian-*.deb.
SHA256SUMS.txt in every release lets you verify the download.
Build from source
git clone https://github.com/Undermove/meridian-browser && cd meridian-browser
npm install
npm start # run from source
npm run dist:mac # or dist:win / dist:linux → installers in dist/
Releases are built by GitHub Actions from a v* tag (.github/workflows/build.yml).
Connect Claude Code (MCP)
-
In Meridian open Settings → AI (or the welcome tour) and pick an agent access level.
-
Click Connect next to Claude Code or Codex: Meridian shows the exact command and runs it after you confirm. Or run it yourself, e.g.
claude mcp add -s user meridian -- node /path/to/meridian-browser/mcp/meridian-mcp.js
For Claude Desktop, add the same command / args to claude_desktop_config.json.
-
Ask: "Read my open GitHub tabs and summarize the PRs", "Find the Grafana dashboard I had open yesterday", "Open three tabs comparing these laptops and put them in split view".
Security model
Agents act inside your logged-in browser, so access is explicit and layered:
| Level | What agents can do |
|---|
| Off (default) | Nothing. The control socket isn't even open. |
| Read only | List tabs and spaces, read pages, screenshots, history search. |
| Ask before acting | Everything else too, but every navigate / click / type / tab change needs your confirmation in a native dialog. |
| Full control | Acts without asking. Meant for agents you trust. |
- Running JavaScript in pages is a separate opt-in, off by default. In "ask" mode the dialog shows the full code (code too long to review is refused), and
type_text shows the full text, including whether it will submit.
- Click confirmations show the element's visible text next to its page-supplied label, its link and its form target. After you approve, the element is checked again so it can't move or be covered while the dialog is open. Dialogs default to Deny.
- The assistant's preset buttons (Summarize, …) never get browser tools, and "use browser" is off until you turn it on for a chat.
- MCP clients connect over a unix socket (named pipe on Windows) with a random name and a random bearer token, both new on every start, advertised only after the socket is bound, in a
chmod 600 file. The bridge checks that the socket is owned by the current user. There's no TCP port, so websites can't reach it, and requests carrying an Origin header are rejected.
- Page text given to models is wrapped as untrusted content, with any attempt to close the wrapper tags neutralized, and is read in an isolated world so pages can't fake it. Models are told never to follow instructions found in pages.
- Agent page scripts run in an isolated JS world, so pages can't tamper with them. Agents can only open
http(s) URLs.
- Mission Control gets agent events over a private local channel: a unix socket in a
0700 temp dir (a random named pipe on Windows), a per-session random token in the agent's environment, size- and rate-limited JSON lines, nothing sent back. Claude Code hooks are added to the session's own --settings file as async hooks (they never block the agent) next to the secret guard; Codex gets a per-session notify override. Hooks send only short summaries (file names, command heads, hosts), which Meridian redacts again (tokens, KEY=value, URL credentials, vault values). Prompts, tool inputs and outputs are never sent; the persisted history holds only labels, times and counts.
- The Claude API key is encrypted with the OS keychain (safeStorage) and only used in the main process. Web content and the UI page never see it.
- Electron hardening: sandboxed UI renderer with context isolation,
<webview> preferences forced safe in will-attach-webview, IPC sender validation, page-initiated navigation to file: or other privileged schemes blocked, server redirects to such schemes blocked, external app links (mailto:, zoommtg: …) allowlisted and confirmed every time, executable and HTML downloads and download bursts need confirmation, WebUSB/HID/Serial and Bluetooth pickers denied, client certificates chosen by you, camera and mic as separate per-site prompts (resettable in Settings), favicons fetched through the site's profile and shown as URLs, dev/test hooks disabled in packaged builds, and Electron fuses (no , no , no inspector flags, ASAR integrity, cookie encryption, no extra privileges).
Found a vulnerability? Please open a private security advisory on GitHub.
Keyboard shortcuts
| macOS | Windows / Linux |
|---|
| New tab / command bar | ⌘T | Ctrl+T |
| Edit URL | ⌘L | Ctrl+L |
| Close tab | ⌘W | Ctrl+W |
| Reopen closed tab | ⇧⌘T | Ctrl+Shift+T |
| Pin / unpin | ⌘D | Ctrl+D |
| Toggle sidebar | ⌘S | Ctrl+S |
| Ask Claude | ⌘J | Ctrl+J |
| Mission Control | ⇧⌘M | Ctrl+Shift+M |
| Split view / close | ⌃⇧= / ⌃⇧- | Ctrl+Shift+= / Ctrl+Shift+- |
| Spaces | ⌃1…9, ⌘⌥←/→ | Alt+1…9, Ctrl+Alt+←/→ |
| Tabs | ⌘1…9, ⌃Tab | Ctrl+1…9, Ctrl+Tab |
| Copy URL | ⇧⌘C | Ctrl+Shift+C |
| New folder | ⌥⇧⌘N | Ctrl+Alt+Shift+N |
| Little window | ⌘⌥N | Ctrl+Alt+N |
| Private window | ⇧⌘N | Ctrl+Shift+N |
| Archive / Downloads | ⌥⌘A / ⇧⌘J | Ctrl+Alt+A / Ctrl+Shift+J |
| Settings / all shortcuts | ⌘, / ⌘/ | Ctrl+, / Ctrl+/ |
Every shortcut can be changed in Settings → Shortcuts.
Known limitations
- Arc logins (cookies) transfer on macOS only. Arc for Windows encrypts them in a way third-party apps can't read (App-Bound Encryption), so there you sign in once per profile. Saved passwords never transfer: export them from Arc into a password manager. Partitioned (CHIPS) cookies are skipped, and session cookies last until you quit Meridian.
- No built-in password manager or sync yet. Use your password manager's desktop app.
- Chrome extensions are experimental: background and content scripts load, but toolbar popups don't yet.
- No Boosts, Easels or Notes.
- Codex has no per-tool hooks, so the secret-file guard and Mission Control's tool view only cover Claude Code. For Codex sessions Meridian falls back to scanning the project for secret files.
- Inside a Docker sandbox git is read-only, by design. Commit from a host terminal.
- Windows and Linux builds are produced by CI but have had far less hands-on testing than macOS.
Roadmap
Extension toolbar, a password manager integration, sync between devices, Boosts-style site tweaks.
Privacy
No account, no server, no telemetry. Everything stays on your computer; AI features talk only to the AI you connect. Details: privacy page.
License
MIT. Third-party components and their licenses: THIRD_PARTY_NOTICES.md. Meridian is an independent project, not affiliated with or endorsed by The Browser Company. "Arc" is a trademark of its respective owner and is used here only to describe compatibility.
