MCP Web Agent (Community)

unlisted

by Edward Shamosh

Lets Claude (or any AI tool) control a browser tab inside Obsidian, using websites you're already logged into. Pairs with the obsidian-mcp-web-agent npm package.

Updated 22d agoMIT

View on GitHub

Obsidian MCP Web Agent — Community Plugin

An Obsidian plugin (plus a small AI connector that runs alongside it) that lets Claude or any AI tool control a browser tab inside Obsidian, using websites you're already logged into.

Sign into Gmail once. Then ask Claude to read your inbox, draft replies, archive newsletters. It happens inside Obsidian, in the panel you already use, with your real session. No password handed to the AI, no separate browser opening up, no copy-pasting between windows.

Demo screenshot goes here. Capture the 🟢 status bar plus the Web viewer mid-task. Save as docs/demo.png and replace this line with ![demo](docs/demo.png).

What you'd use it for

Manage Gmail without giving Claude your password. Sign in once, ask it to triage, draft, archive, search.
Drive any logged-in dashboard. Stripe, Vercel, GitHub, Upwork, Shopify Admin. Claude pulls data, fills forms, runs reports.
Read an article, write the note. Claude opens the page in your Web viewer, reads it, writes a summary straight into your vault.
Test your own web apps. Localhost dashboards, internal tools, prototypes. Claude navigates while you watch.

If you've used Playwright MCP, Chrome DevTools MCP, or agent-browser before, this fills a different slot: the browser that already has all your logins in it.

How it's built (so you know what you're installing)

Two small pieces:

An Obsidian plugin — flips on a "remote control" feature that's already built into Obsidian's browser
A small AI connector (an MCP server) — uses that remote control to drive the browser on Claude's behalf

You install both. They work together. Setup is about 5 minutes total.

Install

What you'll need first

Obsidian (desktop, v1.5.0+)
Node.js v18 or higher (node -v to check; install from nodejs.org if missing)
Git (git --version to check)
Claude Code or Claude Desktop

1. Install the Obsidian plugin (via BRAT)

In Obsidian:

Settings → Community plugins → Browse, search for "BRAT", install + enable
Settings → BRAT → "Add Beta plugin", paste Zbrooklyn/obsidian-mcp-web-agent-community-plugin, click Add Plugin
Settings → Community plugins, toggle "MCP Web Agent" ON

A notice will pop up: "patched N Obsidian shortcut(s). Active on next launch." The plugin just added a launch flag to your Obsidian shortcuts. Quit Obsidian and reopen it normally. Status bar (bottom-right) should show 🟢 Bridge active.

2. Install the AI connector (via npm)

In a terminal:

npm install -g github:Zbrooklyn/obsidian-mcp-web-agent-community-plugin

The connector lives in the /server/ subfolder of the same repo. The npm command pulls and installs from there automatically.

3. Tell Claude about it

If you use Claude Code (terminal):

claude mcp add obsidian-mcp-web-agent -- obsidian-mcp-web-agent

Then in any active Claude Code session, run /reload-plugins. Tools appear instantly.

If you use Claude Desktop (GUI), edit your config file:

OS	Path
Mac	`~/Library/Application Support/Claude/claude_desktop_config.json`
Windows	`%APPDATA%\Claude\claude_desktop_config.json`

Add this under "mcpServers":

{
  "mcpServers": {
    "obsidian-mcp-web-agent": {
      "command": "obsidian-mcp-web-agent"
    }
  }
}

Save, fully quit and reopen Claude Desktop.

4. Try it

Ask Claude:

Use obsidian_status to check the bridge.

You should get back reachable: true and your Chrome version. Then try something real:

Open https://news.ycombinator.com in my Obsidian Web viewer and tell me the top 3 headlines.

You'll watch Claude drive the tab live.

Need a click-by-click walkthrough? See INSTALL.md.

FAQ

Does this work with Claude Desktop? Yes. Same connector, different config file (see install step 3 above).

Is my data safe? Everything runs on your computer. The remote-control port only listens on 127.0.0.1 (your machine, not the network). Cookies stay where they always were, in Obsidian's local data folder. Nothing leaves your machine.

Why not just use Playwright MCP or Chrome DevTools MCP? Those spawn a fresh browser. Different cookies, different logins, doesn't reuse anything you've already authenticated with. The whole point of this is reusing your Obsidian session. If you don't need that, the others are great. (And you can run them all at once. They don't conflict.)

Will this slow down Obsidian? The plugin is small and barely runs. The Web viewer itself can feel heavy when loading sites like Gmail, but that's the site's weight, not the plugin's.

The plugin isn't in the official Obsidian community store. Is it official? Not yet. This is a beta release distributed through BRAT. If usage and stability hold up, the official store comes next.

How it actually works (technical)

Obsidian is an Electron app. Its Web viewer is a Chromium webview. When Electron is launched with --remote-debugging-port=9222, that Chromium exposes the Chrome DevTools Protocol (CDP) on localhost:9222. Anything that speaks CDP — Playwright, Chrome DevTools MCP, raw chrome-remote-interface, this project's MCP server — can attach and drive it.

This plugin's job is making the launch flag a one-toggle setting:

Enable the plugin
It scans your Obsidian shortcuts (Desktop, Start Menu, Taskbar pin) and adds --remote-debugging-port=9222 to each, saving the original args so the toggle can revert
Your next normal launch opens the debug port automatically
Toggle the plugin off and the shortcuts revert cleanly

The plugin is about 600 lines of vanilla JavaScript, no build step. The MCP server is about 900 lines, also vanilla, talks raw CDP over WebSocket. Both pieces are MIT-licensed and reviewable.

MCP tools (25)

The MCP server exposes these to Claude:

Diagnostic: obsidian_status, obsidian_list_tabs, obsidian_list_targets
Tab management: obsidian_new_tab, obsidian_close_tab, obsidian_navigate, obsidian_back, obsidian_forward, obsidian_reload
Reading state: obsidian_read_state, obsidian_read_text, obsidian_get_html, obsidian_snapshot (structured interactive elements list), obsidian_evaluate
Interaction: obsidian_click, obsidian_type (works with React-controlled inputs), obsidian_press_key, obsidian_hover, obsidian_select_option, obsidian_wait_for, obsidian_scroll
Visual / debug: obsidian_screenshot, obsidian_console_messages, obsidian_wait_for_network_idle, obsidian_cookies (names + flags only, values redacted)

Settings, commands, and the status bar

Plugin settings (Obsidian → Settings → MCP Web Agent)

Bridge enabled — master toggle. ON patches your shortcuts. OFF reverts them.
Debug port — defaults to 9222. Change only if it conflicts with another tool.
Patched shortcuts — list of .lnk files the plugin modified, with original args saved.
Restart Obsidian to apply bridge — quits and relaunches via a patched shortcut.

Status bar indicator

State	Meaning
🟢 Bridge active	Debug port is live, AI agents can attach
🟡 Bridge active on next launch	Shortcuts are patched, but THIS Obsidian instance was launched without the flag. Restart.
🔴 Bridge needs setup	Toggle "Bridge enabled" on.
⚪ Bridge disabled	You toggled it off.

Commands (Cmd/Ctrl-Shift-P → "MCP Web Agent:")

Open URL in Web viewer
List CDP targets
Check bridge status now
Restart Obsidian to apply bridge
Enable / Disable bridge

Comparison vs other browser MCPs

Tool	What it drives	When to use
Obsidian MCP Web Agent (this)	Obsidian's Web viewer (your real session)	Logged-in personal sessions; results land in your vault
chrome-devtools-mcp	Fresh Chromium spawned by the MCP	Anonymous browsing, parallel contexts
@playwright/mcp	Playwright-managed browsers	Headless mode, cross-browser testing
browsermcp	Your actual desktop Chrome via extension	Driving the browser you already use

These complement each other. Routing pattern: this for personal logged-in stuff, the others for fresh anonymous work.

Platforms and known limits

Platforms

Windows — fully supported, tested on Win11
macOS — bridge mechanism works, auto-patch step is Windows-specific. Mac users launch Obsidian with open -a Obsidian.app --args --remote-debugging-port=9222. PRs welcome.
Linux — same as macOS. Manual launch with obsidian --remote-debugging-port=9222.
Mobile — not supported. Obsidian Mobile isn't Electron, no debug port.

Privacy

Plugin runs entirely locally. No telemetry.
Debug port binds to 127.0.0.1 only.
Patched shortcut paths are stored in data.json inside the plugin folder. Gitignored, never leaves your machine.
The MCP server only connects to localhost:9222.

Known issues

Cloudflare error 1002 on some sites — Cloudflare refuses to load some sites in Obsidian's webview because of how its DNS resolution looks. Use a regular browser MCP for those specific sites.

Troubleshooting

Status bar stays 🔴 after enabling. The plugin couldn't find shortcuts. Settings → MCP Web Agent → "Patched shortcuts" should list at least one. If empty, run "Patch shortcuts now" from the command palette.

Status bar stays 🟡 forever. You launched Obsidian via a shortcut the plugin didn't patch. Either run "Restart Obsidian to apply bridge" from the command palette, or close + reopen Obsidian via a patched shortcut.

External tool can't see the webview. Open a Web viewer pane in Obsidian first (Cmd palette → "Web viewer: Open"). The plugin doesn't auto-open one.

Tools don't appear in Claude after install. For Claude Code, run claude mcp list to confirm obsidian-mcp-web-agent is registered, then /reload-plugins. For Claude Desktop, check claude_desktop_config.json for syntax errors (no trailing commas) and fully quit + reopen.

Roadmap

Planned:

Cross-platform shortcut patching (macOS, Linux)
Submission to the official Obsidian community store
npm registry publish (currently install via github: URL)
Demo videos

Maybe (open to discussion):

Per-tab visual badge in Obsidian showing which Web viewer Claude is driving
Webhook subscription so external tools can react to navigations

Not in scope:

Replacing your actual browser
Mobile support
Headless mode

Contributing

Issues and PRs welcome at https://github.com/Zbrooklyn/obsidian-mcp-web-agent-community-plugin/issues.

For PRs: open an issue first to discuss. Match the existing code style (vanilla JS, no build step, idiomatic Obsidian Plugin API). Test on your own vault. Update the README and INSTALL.md if your change affects user-facing behavior.

For bug reports, include: Obsidian version, plugin version, OS, what you ran, what you expected, what happened, and the status bar state at the time.

Acknowledgments

Built on:

Obsidian and its plugin API
The Model Context Protocol and the @modelcontextprotocol/sdk
BRAT for beta plugin distribution
The Chrome DevTools Protocol

Inspired by other "AI inside Obsidian" plugins, each doing a different piece:

Claudian (Claude Code as sidebar chat)
obsidian-mind (vault as persistent memory for AI agents)
obsidian-agent-client (Agent Client Protocol bridge)

This project's contribution: the browser bridge specifically.

License

MIT — see LICENSE.

For plugin developers

Search results and similarity scores are powered by semantic analysis of your plugin's README. If your plugin isn't appearing for searches you'd expect, try updating your README to clearly describe your plugin's purpose, features, and use cases.

MCP Web Agent (Community)

Obsidian MCP Web Agent — Community Plugin

What you'd use it for

How it's built (so you know what you're installing)

Install

What you'll need first

1. Install the Obsidian plugin (via BRAT)

2. Install the AI connector (via npm)

3. Tell Claude about it

4. Try it

FAQ

MCP tools (25)

Plugin settings (Obsidian → Settings → MCP Web Agent)

Status bar indicator

Commands (Cmd/Ctrl-Shift-P → "MCP Web Agent:")

Platforms

Privacy

Known issues

Contributing

Acknowledgments

License

Connect MCP

MCP Connector

Vault MCP Server

MCP Server

Claude Code IDE Bridge

Zenith Bridge

Claude Code IDE

Agent Client

ACP Bridge

Vault as MCP

For plugin developers