How do I connect Web2MD to my AI agent?

Run `npm i -g web2md-mcp-server && web2md-mcp-setup`. It detects which MCP clients you have installed — Claude Code, Codex, Cursor, Claude Desktop — and configures each one, then registers the native messaging host so the server can reach your Chrome. Fully quit and reopen Chrome afterwards.

Do I need to log in or get an API key first?

No. Earlier versions required signing in before setup would run. The setup command now signs itself in through a browser OAuth flow when no credentials exist, so the install is a single line with nothing to paste.

Which MCP clients does it configure?

Claude Code and Codex through their own CLIs, and Claude Desktop and Cursor by merging into their JSON config files. Clients you do not have installed are skipped and reported as skipped, so you can see exactly what happened rather than guessing.

Why does it write an absolute node path instead of just 'npx'?

GUI applications like Claude Desktop and Cursor do not inherit your shell PATH, so a bare `npx web2md-mcp` fails when launched from the dock. The setup writes the resolved absolute path to node and to the server entry point, which is why it works from a GUI launch rather than only from a terminal.

Do I have to restart anything?

Chrome, fully — Cmd+Q on macOS or via the system tray on Windows. Chrome only reads native-messaging host manifests on a cold start, so closing the window is not enough. MCP clients pick up their new config on next launch.

What if a client fails to configure?

The summary lists it under skipped along with the exact manual command to run, for example `claude mcp add web2md -s user -- node /dist/index.js`. Nothing fails silently and you are never left guessing which of four clients did not get set up.

What does this actually give my agent?

Three tools: `agent_convert` for one URL, `agent_batch_convert` for up to 50, and `semantic_search` across what you have already converted. All of them read pages through your real signed-in Chrome, which is how they reach Reddit, X trending, and other sites that block server-side crawlers.

One Command Connects Web2MD to Claude Code, Cursor, Codex, and Claude Desktop

MCP setup has a tedious shape. Every client stores its config somewhere different, each wants a slightly different format, and all of them need an absolute path that you have to look up first. Do that four times and you have spent twenty minutes on plumbing before your agent has read a single page.

Short answer

One line now handles all four clients: npm i -g web2md-mcp-server && web2md-mcp-setup. It detects which MCP clients are installed, configures each in its own format, registers the native messaging host, and prints a summary of what it did and what it skipped. There is no login step and no path to look up.

What it looks like

npm i -g web2md-mcp-server && web2md-mcp-setup

web2md-mcp-setup — connecting all your Agents…

▸ Registering local Agent bridge (native messaging host)…
✓ Native messaging host installed
  Manifest: ~/Library/Application Support/Google/Chrome/NativeMessagingHosts/com.web2md.agent.json
  Extension ID: ijmgpkkfgpijifldbjafjiapehppcbcn

Restart Chrome (Cmd+Q then reopen) for the changes to take effect.

═══ Setup summary ═══
  ✓ Claude Code
  ✓ Codex
  ✓ Claude Desktop
  ✓ Cursor

The summary is the part that matters. Anything not installed appears under skipped rather than vanishing, so you know whether three clients were configured or four.

The two things that used to break

The login step. Setup previously required credentials before it would run, which meant the "one command" was really a command, a signup, a token paste, and then the command again. It now completes a browser OAuth flow itself when no credentials exist. If you are already signed in, it uses what you have.

The PATH problem. This one is less obvious and bites harder. Claude Desktop and Cursor are GUI applications — launched from the dock, they do not inherit the PATH from your shell. A config that says npx web2md-mcp works when you test it in a terminal and fails silently when you actually use the app. The setup resolves node and the server entry point to absolute paths and writes those instead, which is why it survives a GUI launch.

What your agent gets

| Tool | What it does | |---|---| | agent_convert | One URL → Markdown, through your Chrome | | agent_batch_convert | Up to 50 URLs → Markdown each, sequential | | semantic_search | Search across what you have already converted |

All three read through your real browser session. That is the reason they reach content that server-side crawlers cannot: Reddit threads, Hacker News, X trending and other sites where a datacenter fetch gets a login wall or an empty shell.

After running it

Fully quit Chrome and reopen — Cmd+Q on macOS, system tray on Windows. Closing the window is not enough, because Chrome reads native-messaging host manifests only on a cold start. This is the single most common reason setup appears to have worked and then the tools are not there.

Then ask your agent to convert something:

Convert https://news.ycombinator.com/ and tell me the three
stories with the most comments.

If the tools are wired up, it calls agent_convert and answers from the real page. If not, restart Chrome first — that is almost always the cause.

Let your agent read the sites that block crawlers — measured coverage across Reddit, HN, X, and the failures
Give your agent Reddit access — the Reddit-specific walkthrough

One Command Connects Web2MD to Claude Code, Cursor, Codex, and Claude Desktop