MCP server and CLI tool for interacting with the Internet Archive's Wayback Machine. Supports full CDX search, snapshot content retrieval, screenshot listing, snapshot comparison, and optional authentication for higher SPN2 rate limits.
MCP server and CLI tool for interacting with the Wayback Machine without API keys
pnpm validate # typecheck + lint + build + test + untested-files check (the full CI gate)
pnpm check # typecheck + lint + build only
pnpm build # compile TypeScript to dist/
pnpm test # run unit and integration tests
pnpm test:coverage # run tests with coverage (80% line/branch/function threshold)
pnpm lint # lint with ESLint
pnpm lint:fix # auto-fix lint issues
To run a single test file:
node --test tests/tools/save.unit.test.ts
End-to-end tests hit the live Wayback Machine API and are opt-in:
pnpm test:e2e # sets WAYBACK_LIVE_TESTS=1 internally via turbo
pnpm validate is the gate that must pass before a release. prepublishOnly runs it automatically.
Architecture
src/bin.ts is the entry point. It detects whether it is invoked as a CLI or loaded as an MCP server and routes accordingly.
src/
bin.ts — entry point; dispatches to CLI or MCP server
cli.ts — Commander-based CLI implementation
server.ts — MCP server wiring (ListTools + CallTool handlers)
contexts.ts — shared context (rate limiter, cache, credentials)
schemas.ts — Zod schemas for all tool inputs; single source of truth
tools/
save.ts — save_url tool (SPN2 API)
retrieve.ts — get_archived_url tool
search.ts — search_archives tool (CDX API)
status.ts — check_archive_status tool (sparkline API)
screenshots.ts — list_screenshots tool
compare.ts — compare_snapshots tool
cache.ts — clear_cache tool
context.ts — injects shared context into tool handlers
utils/
http.ts — fetch wrapper with rate limiting and Retry-After handling
cache.ts — in-memory + disk cache with per-endpoint TTLs
rate-limit.ts — 15 req/min token bucket
validation.ts — shared Zod validation helpers
Each tool module exports a schema (consumed by ListToolsRequestSchema) and an execution function (consumed by CallToolRequestSchema). New tools need both registrations in server.ts.
Caching TTLs are intentional — do not normalise them:
Resource
TTL
Reason
Snapshot content
24 h
Immutable once captured
Availability, CDX, sparkline
1 h
Grows but never mutates
Save operations
30 min
Idempotent per URL
Save status polling
30 s
Changes during active jobs
Conventions
TypeScript strict mode with noUncheckedIndexedAccess and exactOptionalPropertyTypes — no any, no as assertions.
ES Modules throughout — "type": "module" in package.json. Always use .ts extensions in relative imports (rewritten to .js at build time via rewriteRelativeImportExtensions).
Zod is the single source of truth for all tool input shapes. schemas.ts defines them; zodToJsonSchema derives the MCP-compatible JSON Schema.
Prettier formats all TypeScript: 4-space indent, double quotes, trailing commas (es5), 80-char print width, LF line endings.
Conventional commits are enforced by commitlint. Allowed scopes: retrieve, save, search, status, fetch, http, validation, cli, build, release, ci, deps. Commit messages must use British English.
Test colocation: unit tests in tests/tools/*.unit.test.ts and tests/utils/*.unit.test.ts; integration tests in tests/*.integration.test.ts. Use the Node.js built-in test runner — no Jest or Vitest.
erasableSyntaxOnly: true — no TypeScript syntax that cannot be stripped without transformation (no enum, no decorators, no namespace).
Gotchas
pnpm validate before pushing. CI runs check + test + coverage + untested-files. pnpm validate replicates this locally via Turbo.
Turbo caches aggressively. If you change a config file that Turbo doesn't track as an input, cached task results may be stale. Clear with pnpm turbo run <task> --force if results look wrong.
WAYBACK_LIVE_TESTS must be set to run test:e2e. The Turbo config passes it through via globalPassThroughEnv; don't set it in .env files — export it in your shell before running.
Coverage excludessrc/contexts.ts, src/cli.ts, src/bin.ts, and src/tools/context.ts — these are wiring/entry-point files. The 80% threshold applies to the remaining surface.
Rate limiting is 15 req/min across all Wayback Machine API calls, with automatic Retry-After handling for 429 responses. Tests that mock HTTP must respect this — don't call real endpoints from unit tests.
noUncheckedIndexedAccess means Record<string, T> lookups return T | undefined. Never fall back to ?? default — narrow explicitly or restructure to a concrete type.
Node version is pinned in .tool-versions. CI tests against Node 22, 24, and 26. Do not use Node APIs that aren't available in 22.
Contributing
Commits follow Conventional Commits and are lint-checked by commitlint on PRs. PRs target main; CI must pass (check, test, coverage). Releases are fully automated via semantic-release on push to main.
After release, alias packages (wayback-machine-mcp, mcp-internet-archive, internet-archive-mcp, @mearman/mcp-wayback-machine) are published automatically by CI — do not publish these manually.
Installation
As an MCP server
CLI shorthand
Claude Code (MCP):
claude mcp add wayback-machine -- npx -y mcp-wayback-machine
The env block is optional — the server works anonymously without credentials.
As a CLI tool
npx mcp-wayback-machine save https://example.com
Or install globally:
npm install -g mcp-wayback-machine
wayback save https://example.com
As a Cloudflare Worker
Deploy the MCP server as a stateless Cloudflare Worker. Runs on the free tier with no paid bindings — all persistent state uses the Cache API which has no published daily limits.
pnpm add -D wrangler
wrangler deploy
The Worker uses the SDK's StreamableHTTPServerTransport in stateless mode (no session IDs), so each request is independent and cold starts are handled gracefully.
Environment variables (set via wrangler secret put):
Variable
Required
Purpose
WAYBACK_ACCESS_KEY
No
Fallback IA S3 credentials for higher SPN2 rate limits
WAYBACK_SECRET_KEY
No
Fallback IA S3 credentials
MCP_AUTH_TOKEN
No
Bearer token for client authentication
When MCP_AUTH_TOKEN is set, clients must send Authorization: Bearer <token>. When absent, the Worker accepts unauthenticated requests.
Per-request credentials. Clients can pass their own IA S3 credentials on each request via HTTP headers, overriding the server's environment variables:
This lets multiple users share a single Worker deployment while each using their own credentials for higher SPN2 rate limits. When both headers are present, they take precedence over WAYBACK_ACCESS_KEY/WAYBACK_SECRET_KEY. When absent, the Worker falls back to its environment variables.
Worker-specific files are excluded from the main tsconfig.json and type-checked separately via tsconfig.worker.json (which adds @cloudflare/workers-types).
Architecture. The stdio and Worker deployments share the same tool logic through pluggable interfaces:
Component
Stdio
Worker
Caching
DiskCacheBackend (OS cache dir)
CacheApiBackend (caches.open())
Rate limiting
InMemoryRateLimiter
CacheApiRateLimiter
Auth
None
StaticTokenAuthProvider (optional)
Credentials
Environment variables
Request headers, then environment variables
Quick examples
Archive https://example.com to the Wayback Machine
Find all archived snapshots of https://example.com from 2023
What's the earliest archived version of https://example.com?
Compare the oldest and newest snapshots of https://example.com
Check how many times https://example.com has been archived
Tools
save_url
Archive a URL to the Wayback Machine using the SPN2 API.
Parameter
Required
Description
url
Yes
The URL to archive
captureScreenshot
No
Capture a screenshot as a PNG image
captureOutlinks
No
Also archive up to 100 outlinked pages
ifNotArchivedWithin
No
Skip if archived within timeframe, e.g. "30d"
jsBehaviorTimeout
No
Run JavaScript for N seconds before capturing (max 30)
forceGet
No
Use simple HTTP GET instead of browser rendering
delayWbAvailability
No
Delay indexing ~12 hours to reduce server load
get_archived_url
Retrieve an archived snapshot's content and metadata.