Overview

The Wayback Machine and Internet Archive library (40M+ items). Find and fetch archived snapshots of any URL, search the library by keyword and metadata, and retrieve item metadata, file manifests, and OCR text.

Tools

Resources

All resource data is also reachable via ia_get_item.

Capability reference

ia_find_snapshots _tool

• closest mode: returns the nearest capture to a given timestamp via the Availability API, falling back to one CDX closest-capture query (preferring a 200 capture, 25 s deadline) when the Availability API has no answer
• history mode: full capture list via the CDX API; filter by date range (from/to) and HTTP status (status_filter)
• For an exact URL with timestamp:N collapse, pass the returned resume_key with the same URL to continue without skipping the page boundary. The cursor expires after one hour and retains the original filters; limit may change between pages. An invalid cursor returns invalid_resume_key; conflicting explicit filters return resume_key_conflict. Legacy CDX keys and wildcard histories remain available with a notice that pages may omit captures
• collapse is timestamp:8 (one capture per day) when omitted or blank; adjustable to timestamp:N, N=1–14, with timestamp:14 listing one row per distinct capture second
• Up to 10,000 records per call (limit, default 100); a larger limit (up to 150,000) is served as 10,000, setting truncated and a notice naming resume_key when captures remain; resume_key pagination for large histories
• Replay URLs are always https://web.archive.org/web/…
• Typed errors: missing_timestamp (closest mode without a timestamp, answered before any lookup), no_snapshots (no matches), no_snapshot_available (closest mode, no capture near timestamp from either source; or, in either mode, the Wayback Machine does not serve the URL — a 403, 410, or 451, as for a site excluded from it), cdx_unavailable (CDX 5xx, 429, another 4xx, timeout, network error, or unreadable response, or the closest-mode CDX check did not complete), availability_unavailable (closest mode, Availability API 5xx, 429, another 4xx, timeout, network error, or unreadable response); a 429, another 4xx, or a request timeout is answered after one request, a 429 with a hint to wait and a 4xx with a hint that repeating the request will not help

ia_get_snapshot _tool

• Resolves to the nearest available capture when the exact timestamp has no snapshot; exact 14-digit timestamps skip resolution
• Reports the capture Wayback served — replay_url, resolved_timestamp, and resolved_status follow Wayback's redirect when the requested timestamp is not itself a capture
• Decodes the page with its declared charset (Content-Type, then <meta>, else UTF-8 when the bytes are valid UTF-8, else Wayback's guessed charset or windows-1252), then removes scripts, styles, comments, and tags and decodes character references, returning readable plain text alongside the replay URL
• Reads at most the first 4 MiB of a page (a notice says when a page is longer); output capped at IA_MAX_SNAPSHOT_CHARS (default 50,000 characters)
• Typed errors: no_snapshot_available (no capture near the timestamp, the capture is an archived error page — the message names its timestamp and archived status — or Wayback refused the replay with a 4xx, or the lookup with a 403, 410, or 451, as it does for a site excluded from the Wayback Machine), content_fetch_failed (Wayback unreachable, rate-limiting, refusing the lookup, or timed out during lookup or fetch)

ia_search_items _tool

• Solr query syntax plus structured filters: mediatype, collection, creator, language, and date range (date_from/date_to)
• Date bounds accept YYYY, YYYY-MM, or YYYY-MM-DD: date_from uses the period's first day and date_to its last day, including leap years. Blank bounds leave that side open; invalid dates return invalid_date_range before a search request
• mediatype takes the ten Internet Archive media types — texts, movies, audio, software, image, data, web, collection, etree, account — case-insensitively, and resolves common near-misses (text, book, books → texts; movie, video, videos → movies; images → image; collections → collection)
• Sort by downloads, date, title, or random (sort, Solr syntax such as date asc or titleSorter asc; default downloads desc; Advanced Search has no relevance sort). Ties under the sort are ordered by identifier, so consecutive pages neither repeat nor skip an item
• query is at most 2,047 characters, the longest Advanced Search accepts; the filters count toward it, and a URL in the query must be quoted
• Up to 200 results per call (rows, default 50), 1-indexed page; on page 1 a larger rows (up to 10,000) is served as 200, setting truncated and a notice naming rows: 200, page: 2 when more results remain
• Advanced Search serves only the first 10,000 results, so page × rows ≤ 10,000
• Output carries total_found, page, and the rows applied for pagination; an empty page returns a notice rather than an error — naming the mediatype applied when nothing matched, or the last page when page is past the end
• title, date, creator, and collection may be a string or a list; collection leaves out per-account fav- favorites lists, and creator lists the first 10 names, with creator_count when a record has more
• Typed errors answered before any search request: invalid_date_range (an invalid date bound), invalid_mediatype (any other mediatype, listing the accepted values), rows_over_cap (rows above 200 past page 1), and page_limit_exceeded (page × rows past 10,000, naming the last reachable page)
• Typed errors from Advanced Search: invalid_query (it rejected the query syntax, an unquoted URL, a query its filters take past 2,047 characters, or the sort field; its message is relayed) and search_unavailable (it failed, rate-limited the request, sent an unreadable response, timed out, or could not be reached); a 5xx or network error is retried, while a 429, another 4xx, a timeout, or an error body is answered after one request, a 4xx with a hint that repeating the request will not help

ia_get_item _tool

• Returns title, creator, description, subject, collection, licenseurl, rights, and language when present in upstream metadata; creator, description, subject, collection, and language may be a string or a list
• files[] is one page of the manifest in upstream order — format, size, md5, and a direct download_url per file; file_count is always the full manifest size
• max_files (default 50, at most 500 per call) and file_offset (default 0) page through large items; a larger max_files (up to 300,000) is served as 500. When files remain, the response sets truncated and names the next file_offset
• format keeps one file type (exact match, case-insensitive — e.g. DjVuTXT, Text PDF, VBR MP3) before paging and reports the match count as totalCount; a format with no matches returns an empty page and a notice listing the formats the item has
• Typed errors: item_not_found (unknown or dark identifiers), archive_unavailable (the Metadata API failed, rate-limited or refused the request, timed out, could not be reached, or sent an unreadable response); a first timeout is retried once, so a call waits at most two request timeouts

ia_get_text _tool

• max_chars and char_offset page through long documents; has_more signals additional text remains. max_chars defaults to and is capped at IA_MAX_SNAPSHOT_CHARS — a larger value is served at that maximum, with a notice, and the output max_chars reports the value applied
• For sequential pages, pass next_continuation back as continuation with the same identifier, omitting char_offset. The cursor expires after one hour, binds the page to the selected text file, and resumes from its byte checkpoint. If a stable file validator or cursor storage is unavailable, the response gives an offset continuation notice
• Continuation errors: ambiguous_pagination when both token and offset are supplied, invalid_continuation for an invalid or changed-file cursor, and range_unavailable when the download cannot resume at the saved range. Recovery names the char_offset to use
• The upstream read stops at the page's end, so memory per call is bounded by the page size
• Locates the best available text file — DjVuTXT preferred, falls back to plain text; source_file names the file fetched and source_file_bytes its size in bytes, from the item manifest
• Typed errors: item_not_found, no_text_file, download_forbidden (restricted collections), archive_unavailable (the Metadata API or the text download failed, rate-limited or refused the request, timed out, or could not be reached); the metadata request and the download each retry a first timeout once and stop at a second

ia://item/{identifier} _resource

• Returns application/json — title, creator, mediatype, description, subject, collection, date, licenseurl, rights, language, and file_count
• identifier comes from ia_search_items results
• Typed errors: item_not_found for unknown identifiers, archive_unavailable when the Metadata API fails

Features

Built on @cyanheads/mcp-ts-core: stdio and Streamable HTTP transports, pluggable auth (none / jwt / oauth), swappable storage (in-memory, filesystem, Supabase, Cloudflare KV/R2/D1), structured logging with optional OpenTelemetry tracing.

Internet Archive-specific:

• No credentials required — all four APIs are public
• Three service layers: WaybackService (Availability + CDX), ArchiveSearchService (Solr), ArchiveMetadataService (Metadata + downloads)
• CDX collapse-by-day default and configurable limit keep responses tractable for high-capture URLs
• Identifies via a custom User-Agent on every request as required by IA's terms of use; configurable via IA_USER_AGENT

Agent-friendly output:

• Pagination context on every list response — total_found, page, rows (search), resume_key (CDX history), and file_count plus the next file_offset (item files) so agents never have to guess whether results are complete
• Typed error reasons with recovery hints so callers can retry, correct a filter, or recover an expired continuation without parsing error messages
• Structured file manifests — ia_get_item returns file-level metadata (format, size, URL) and a format filter, so agents can pick the right file without paging through thumbnails

Getting started

Public Hosted Instance

A public instance is available at https://internet-archive.caseyjhand.com/mcp — no installation required. It serves Streamable HTTP; each client takes its own configuration shape.

For Claude Code:

claude mcp add --transport http internet-archive-mcp-server https://internet-archive.caseyjhand.com/mcp

For Cline (cline_mcp_settings.json):

{
  "mcpServers": {
    "internet-archive-mcp-server": {
      "type": "streamableHttp",
      "url": "https://internet-archive.caseyjhand.com/mcp"
    }
  }
}

For Cursor (mcp.json):

{
  "mcpServers": {
    "internet-archive-mcp-server": {
      "url": "https://internet-archive.caseyjhand.com/mcp"
    }
  }
}

Claude Desktop adds remote servers under Settings → Connectors.

Self-Hosted / Local

No API key required — the Internet Archive's APIs are fully public.

Add the following to your MCP client configuration file:

{
  "mcpServers": {
    "internet-archive-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/internet-archive-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with npx (no Bun required):

{
  "mcpServers": {
    "internet-archive-mcp-server": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@cyanheads/internet-archive-mcp-server@latest"],
      "env": {
        "MCP_TRANSPORT_TYPE": "stdio",
        "MCP_LOG_LEVEL": "info"
      }
    }
  }
}

Or with Docker:

{
  "mcpServers": {
    "internet-archive-mcp-server": {
      "type": "stdio",
      "command": "docker",
      "args": [
        "run", "-i", "--rm",
        "-e", "MCP_TRANSPORT_TYPE=stdio",
        "ghcr.io/cyanheads/internet-archive-mcp-server:latest"
      ]
    }
  }
}

For Streamable HTTP, set the transport and start the server:

MCP_TRANSPORT_TYPE=http MCP_HTTP_PORT=3010 bun run start:http
# Server listens at http://localhost:3010/mcp

Prerequisites

• Bun v1.4.0 or higher (or Node.js v24+).
• No external accounts or API keys required.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/internet-archive-mcp-server.git

2. Navigate into the directory:

cd internet-archive-mcp-server

3. Install dependencies:

bun install

4. Configure environment:

cp .env.example .env
# Optional: edit .env for custom User-Agent, timeouts, etc.

Configuration

All configuration is validated at startup via Zod schemas in src/config/server-config.ts.

See .env.example for the full list of optional overrides.

Running the server

Local development

• Build and run:

  shCopy

  # One-time build
  bun run rebuild
  
  # Run the built server
  bun run start:stdio
  # or
  bun run start:http
  

• Run checks and tests:

  shCopy

  bun run devcheck   # Lint, format, typecheck, security
  bun run test       # Vitest test suite
  bun run lint:mcp   # Validate MCP definitions against spec
  

Docker

docker build -t internet-archive-mcp-server .
docker run --rm -p 3010:3010 internet-archive-mcp-server

The Dockerfile defaults to HTTP transport, stateless session mode, and logs to /var/log/internet-archive-mcp-server. OpenTelemetry peer dependencies are installed by default — build with --build-arg OTEL_ENABLED=false to omit them.

Project structure

Development guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic
• Use ctx.log for request-scoped logging, ctx.state for tenant-scoped storage
• Register new tools and resources via the barrels in src/mcp-server/*/definitions/index.ts
• Wrap external API calls: validate raw → normalize to domain type → return output schema; never fabricate missing fields

Contributing

Issues are welcome. Run checks and tests before submitting:

bun run devcheck
bun run test

License

Apache-2.0 — see LICENSE for details.