MCP Everything

Get started · Run your own server · How it works · Security · Docs

Most websites have no API, and protocols like MCP only help once a site joins in. MCP Everything does not wait for the site: it works on any website you can use in Chrome. You record a task once, and everyone on your server gets it as a tool in Claude. Each person's logins stay their own, encrypted.

Get started

You need Chrome and an invite from the admin of a server. Want your own server? See Run your own server.

1. Connect Claude

In claude.ai or Claude Desktop: Settings → Connectors → Add custom connector, with https://<server>/mcp. In Claude Code:

claude mcp add --transport http mcp-everything https://<server>/mcp

Then type /mcp and choose Authenticate. Claude opens your server's login page. New here? Paste your invite and pick a name and a password.

2. Install the extension

The welcome page offers the download. Unzip it, open chrome://extensions, turn on Developer mode, click Load unpacked and choose the folder mcp-everything-extension. The page connects it to your account by itself.

3. Record a site

Open the site, log in if it has a login, and click Start recording in the extension. The first time, Chrome asks for access to the site. Do the task once, then click Stop and make tools. That takes about a minute; you can close the popup meanwhile.

4. Link your account on that site

Tools are shared, logins are not. For a site with cookies, click Link session in the extension (a recording does this for you). For a site with a password or its own OAuth, ask Claude to connect: you get a one-time page on your server, and your password never goes through the chat.

5. Ask Claude

Ask in plain words. A tool that changes something runs once, even if Claude retries. A tool that can't be undone, like placing a bid, asks you first, with the exact values it will send. New tools show up in a new chat, or after /mcp → Reconnect in Claude Code.

All the details, like other ways to log in and how to log out, are in docs/using.md.

Related MCP server: Web2Actions

Run your own server

You need Docker and an Anthropic API key.

cp .env.example .env
openssl rand -base64 32    # paste as CONNECTOR_SECRET_KEY in .env, and add your ANTHROPIC_API_KEY
docker compose up -d --build
docker compose exec connector python -m connector.users invite --uses 3

Keep a copy of CONNECTOR_SECRET_KEY: without it, the vault does not open. For claude.ai, put the server behind a reverse proxy with TLS and set PUBLIC_URL. Configuration, limits and the review of tools from guests are in docs/server.md.

How it works

flowchart LR
    subgraph you ["👤#nbsp; You"]
        direction TB
        A(["🤖#nbsp; Claude<br/>or any MCP client"])
        U(["🧩#nbsp; Chrome + extension"])
    end
    subgraph server ["🖥️#nbsp; Your server"]
        direction TB
        P["⚙️#nbsp; Pipeline<br/>code + one Claude call"]
        R[("📚#nbsp; Registry<br/>tools for every user")]
        V[("🔐#nbsp; Vault<br/>logins per user")]
        S["🔌#nbsp; MCP endpoint"]
    end
    subgraph web ["🌐#nbsp; The website"]
        direction TB
        W1["Plain HTTP"]
        W2["Headless Chrome<br/>for bot-protected sites"]
    end
    U -- "① record" --> P
    P -- "② ③ tools" --> R
    U -. "④ your login" .-> V
    A -- "⑤ MCP + OAuth" --> S
    R --> S
    V --> S
    S --> W1
    S --> W2

    classDef plain fill:#8b93b81a,stroke:#8b93b8
    classDef ai fill:#d977571f,stroke:#d97757
    classDef store fill:#5a6ff02a,stroke:#5a6ff0
    class U,P,S,W1,W2 plain
    class A ai
    class R,V store
    style you fill:#d977570d,stroke:#d9775766
    style server fill:#5a6ff00d,stroke:#5a6ff066
    style web fill:#2f9e7a0d,stroke:#2f9e7a66
    linkStyle default stroke:#8b93b8,stroke-width:1.5px

The model never sees a cookie or a password, and never writes a request. Running a tool uses no model at all.

A tool is data, not code: a JSON manifest with the request, its inputs and how to read the answer (docs/format.md). It runs as a plain HTTP request, or as a fetch in a headless Chrome for sites with bot protection (docs/browser-runtime.md).

Security

• Secrets never reach the model or the shared tools. Your sessions go to your own vault, encrypted with AES-256-GCM.

• You type passwords on your own server, through a one-time link. Never in the chat.

• Claude logs in with OAuth 2.1. Tokens are stored hashed, and refresh tokens are replaced on every use.

• You confirm what can't be undone, bound to the exact values shown.

• Every request goes to the site itself, over HTTPS. No captcha solving and no getting around a block.

• Invite only, with limits per person, and an optional review of tools from guests.

The full model and our principles are in docs/security.md.

FAQ

Does Claude ever see my password? No. You type it on a page of your own server, and it only goes to that site.

Can other users see my sessions? No. Tools are shared, sessions are not: a tool always runs with the session of the person who calls it.

How many LLM calls does it take? One per recording, to choose and name the tools. Calling a tool uses no model.

Is this a scraper? No. It replays the requests a site's own frontend made, on your own account, one call at a time. It does not crawl, and it stops when a site blocks it.

More questions in docs/faq.md.

Documentation

Roadmap

• The extension as a runtime: run tools in your own logged-in tab

• Log in with Google or GitHub

• Repair: a new recording replaces a broken tool

• Test a larger set of web shops

License

Proprietary. All rights reserved; see LICENSE.

Built for agents. Polite to websites.

Get started · Run your own server · Tool format · Roadmap