mcp-libgen

MCP server that searches Library Genesis for EPUBs, stores them in Google Drive, and emails them to a Kindle. Runs on Vercel so it works from the Claude mobile app, not just a local desktop client.

How it works

search_libgen(query)      HTTP only. Fetches the search page, extracts the id
                          list behind its JSON tab, resolves metadata through
                          json.php. No browser automation.
        |
        v   model picks a row
download_book(md5)        Fetches the file, validates it is a real EPUB,
                          uploads it to Google Drive.
        |
        v
send_to_kindle(file_id)   Pulls it back out of Drive and emails it.

Drive sits between the two halves because they are separate Vercel invocations on potentially different instances, and /tmp is not guaranteed to survive between them. It also gives you a durable library, so list_library can re-send a book without touching libgen again.

No Playwright. Serverless bundles cap out around 250 MB and Chromium does not fit. Everything here is plain fetch.

Related MCP server: mcp-google-drive

Setup

1. Install

npm install

2. Google Drive credentials

Service accounts do not work here. They have no Drive storage quota on personal (non-Workspace) Google accounts, so they cannot own files in your My Drive. You need OAuth2 with a stored refresh token.

1. Google Cloud Console, new project, enable the Google Drive API.

2. Credentials -> Create credentials -> OAuth client ID. Either application type works. If you pick Web application, you must add http://localhost:53682/callback under Authorised redirect URIs or the consent flow fails with redirect_uri_mismatch. Desktop app clients accept loopback addresses without registering them.

3. OAuth consent screen -> Audience:

   • User type External (personal Gmail cannot use Internal).

   • Add yourself under Test users.

   • Then Publish app to move from Testing to In production. Refresh tokens issued while an app is in Testing expire after 7 days, and the resulting failure days later does not obviously point back at this.

4. Download credentials.json into the project root. npm run auth reads it directly, so you do not need to copy the id and secret by hand. It is gitignored.

5. Run the one-time consent flow:

npm run auth

Keep that terminal open, approve in the browser, and paste the resulting GOOGLE_REFRESH_TOKEN into .env.local.

The scope is drive.file, which is per-file access to files this app creates. It cannot see the rest of your Drive, and it is not a restricted scope, so publishing the consent screen needs no verification review.

3. Gmail app password

Requires 2FA on the account. Generate at myaccount.google.com/apppasswords and put the 16 characters in GMAIL_APP_PASSWORD, with no spaces. This is not your Google account password.

4. Amazon approved sender

In Amazon -> Manage Your Content and Devices -> Preferences -> Personal Document Settings, add SENDER_EMAIL to the Approved Personal Document E-mail List. It must match exactly. If it does not, Amazon discards the email silently and nothing in this server can detect it.

5. MCP auth token

node -e "console.log(require('crypto').randomBytes(32).toString('hex'))"

Put it in MCP_AUTH_TOKEN. The endpoint downloads files and sends mail from your Gmail, so it must not be open.

6. Verify credentials

npm run check

Exchanges the refresh token, calls the Drive API, and authenticates against Gmail SMTP, without writing a file to Drive or sending an email. Run this before deploying, and again first thing whenever something breaks: most failures in this system are credential problems that surface far from their cause.

7. Deploy

vercel deploy

Add every variable from .env.example to the Vercel project's environment variables. Raise maxDuration in app/api/[transport]/route.ts as far as your plan allows. libgen mirrors are slow, and the cap is a hard kill with no partial result.

8. Smoke test before anything else

curl -H "Authorization: Bearer $MCP_AUTH_TOKEN" "https://<deployment>/selftest?q=of+mice+and+men"

This exercises search and a real download from inside a Vercel function, which answers the one remaining unknown: does libgen serve Vercel's IPs at all? Vercel egresses from AWS ranges, and Cloudflare-fronted mirrors challenge datacenter traffic more than residential. Working from your laptop proves nothing about this. If it fails here, the host choice is wrong and no amount of code fixes it.

Add &download=0 to test search only.

9. Connect to Claude

Add as a custom connector pointing at https://<deployment>/api/mcp, with the Authorization: Bearer <MCP_AUTH_TOKEN> header.

Local commands

npm run check      # validate all credentials, no side effects
npm run preview -- "of mice and men steinbeck"   # see what the model would see
npm run auth       # one-time Google consent flow

Known limits

• No delivery confirmation. Amazon does not acknowledge personal document delivery. send_to_kindle can only report that the message was accepted by the mail server. If a book never appears, the approved-sender list is the first thing to check.

• Mirrors rotate. libgen domains get seized and move. LIBGEN_MIRRORS is a comma-separated list tried in order; update it when one dies. This is the most likely reason for this to stop working months from now.

• 50 MB attachment cap, enforced before sending.

• No language field. Verified against the live API: neither object=f nor object=e returns a language, even though the HTML results table shows one. There is deliberately no language filter, because it would silently match nothing. The model infers language from title and author instead.

• fields=* does nothing. Tested against libgen.li. File records never carry titles, so the second object=e lookup is mandatory, not a fallback.

• Broad queries are truncated. A query like "dune frank herbert" matches 2000 files. Resolving all of them would take twenty-plus sequential round trips and blow the function's time budget, so scanning stops early and the result table says so.

• Metadata is sparse. Edition records frequently have blank author, year, and publisher. Blank columns are dropped from the table, and title/author fall back to parsing the locator shelf path.

• Results routinely include study guides, omnibus editions, and abridgements alongside the actual book. Picking correctly is left to the model on purpose; it is not reliably decidable in code.