Internet Archive MCP Server

Full-coverage Internet Archive MCP server — search, metadata, collections, and Wayback Machine in one server.

Why this server?

Existing Internet Archive MCP servers are almost all Wayback-only — small 3–4 tool wrappers around the CDX snapshot API. This is the only server that combines:

• Full IA collections access — search, metadata, files, reviews, and collection browsing across all media types (texts, audio, movies, software, images)

• Full Wayback Machine coverage — CDX snapshot search, availability checks, and raw content fetch

• 13 tools organized in three tiers (core, differentiator, auth-gated)

• Anonymous reads — every read operation works with zero configuration

• Auth-gated writes — Save Page Now requires IA S3 keys, cleanly separated from the read path

Quick start

Install

cd internet-archive-mcp
pip install -e .

Hermes Agent

Add to your Hermes config.yaml:

mcp_servers:
  internet-archive:
    command: python3
    args: ['-m', 'internet_archive_mcp.server']
    env:
      PYTHONPATH: /path/to/internet-archive-mcp/src

Claude Desktop

Add to claude_desktop_config.json:

{
  "mcpServers": {
    "internet-archive": {
      "command": "python3",
      "args": ["-m", "internet_archive_mcp.server"],
      "env": {"PYTHONPATH": "/path/to/internet-archive-mcp/src"}
    }
  }
}

Tool reference

Tier 1 — Core tools

Tier 2 — Differentiator tools

Tier 3 — Auth-gated tools

Configuration

All read tools work anonymously with no configuration. The only tool requiring credentials is save_page (Save Page Now).

Environment variables

Get your keys at: https://archive.org/account/s3.php (free account required).

Keys can also be passed directly as tool arguments (access_key, secret_key) — env vars are the fallback.

Rate limiting & caching

The client is designed to be a polite, well-behaved consumer of IA APIs:

• Rate limiter — minimum 0.5s interval between all outgoing requests

• Bounded TTL cache — in-memory cache with 256-entry cap; search and metadata results cached for 1 hour; oldest entries evicted first

• 429 retry with backoff — automatic retry on HTTP 429 with exponential backoff (honors Retry-After header), up to 3 attempts

• Mandatory User-Agent — every request carries a descriptive User-Agent: internet-archive-mcp/0.1.0 (...) header per IA policy

Development

# Install with dev dependencies
pip install -e '.[dev]'

# Run tests
pytest tests/ -v

Architecture

internet-archive-mcp/
├── src/internet_archive_mcp/
│   ├── __init__.py
│   ├── client.py      # Async httpx client — all API logic
│   └── server.py      # FastMCP tool definitions (thin wrappers)
├── tests/
│   ├── __init__.py
│   ├── conftest.py
│   ├── test_client.py
│   ├── test_qa_fixes.py
│   ├── test_qa_round2.py
│   ├── test_qa_round3.py
│   ├── test_server.py
│   └── test_tools.py
├── RESEARCH.md        # API research findings (tested with curl)
└── pyproject.toml

• client.py — ArchiveClient class using async httpx. Talks to two base URLs: https://archive.org (search, metadata, availability) and https://web.archive.org (CDX, content fetch, Save Page Now). Owns caching, rate limiting, and retry logic.

• server.py — FastMCP server with 13 @mcp.tool() definitions. Each tool is a thin async wrapper that delegates to ArchiveClient and converts exceptions to error strings.

Design decisions

The Related Items API (archive.org/recommendations/{id}) was investigated during research but returned no useful data. It is intentionally excluded. See RESEARCH.md §5f.

Links

• RESEARCH.md — full API research with tested curl examples

• Internet Archive Developer Docs

• IA Bots Policy — terms of service and acceptable use for automated access