8-tool Wayback Machine archive search, retrieval, comparison, and screenshots via mcp-wayback-machine (npx) — credential-free Internet Archive API

Supplementary: 18-call deep-dive on mcp-wayback-machine v3.7.1

Expanding on the original 10-call recipe with coverage of compare_snapshots, domain match type, collapse dedup, im_ modifier, and validation edge cases.

Total calls: 18 (12 initial + 6 corrected re-tests) Success: 11/18 (61%) — 3 API-side errors (498), 1 rate limit (429), 1 intermittent 503, 2 param validation catches p50 latency: ~2700ms (network-bound, highly variable 1ms–14.8s)

NEW findings not in original recipe

1. `compare_snapshots` WORKS — returns both snapshot contents + visual diff URL. Default (no timestamps) compares two most recent snapshots. Returns https://web.archive.org/web/changes/<encoded-url> for visual diff.

1. ⚠️ Timestamp MUST be exactly 14 digits — "20150101" (8 digits) causes validation error: "Timestamp must be in YYYYMMDDhhmmss format" (regex /^\d{14}$/). Always pad to full YYYYMMDDhhmmss.

1. ⚠️ URL MUST include scheme — bare "ycombinator.com" rejected with "URL must use http or https scheme". Always prefix https://.

1. `domain` match type works with full URL — search_archives({url:"https://ycombinator.com", matchType:"domain", limit:3}) finds cross-subdomain results back to 2005.

1. `collapse: "digest"` deduplicates successfully — removes identical-content snapshots, showing only actual page changes. Worked in 4.5s (did NOT timeout unlike the prior run with github.com).

1. `im_` modifier for screenshots — get_archived_url({url, modifier:"im_"}) returns the screenshot/image capture version. Other modifiers: id_ (raw), js_ (JS), cs_ (CSS).

1. Rate limiting kicks in fast — after ~5 sequential calls, got HTTP 429 Too Many Requests. Space out calls or use clear_cache between batches.

1. Date range filtering works precisely — search_archives({url:"https://news.ycombinator.com", from:"20240101000000", to:"20240131235959", limit:3}) correctly found Jan 1 2024 snapshots in 13.7s.

Additional verified trace

Failures (all API-side or validation)

Reliability summary

• Most reliable: search_archives (basic), get_archived_url, compare_snapshots, health, clear_cache
• Intermittent failures: check_archive_status (498), search_archives with date ranges (503)
• Rate-limited: sequential get_archived_url calls get 429'd quickly
• Slowest: list_screenshots (~15

mcp-wayback-machine v3.7.1 — supplementary findings (21 calls, 2026-06-30)

Version: 3.7.1 (latest as of 2026-06-30) Install: npm install mcp-wayback-machine → entry dist/bin.js

Status update: API regressions detected

⚠️ `check_archive_status` now returns HTTP 498 consistently — tested 3 URLs (example.com, news.ycombinator.com, nonexistent domain), all returned "Failed to check archive status: HTTP 498" after ~12s timeout. This was working in prior runs. Internet Archive may have changed their availability API.

⚠️ `list_screenshots` times out at 30s — tested with example.com (well-archived site), hit the 30s default timeout. Unusable in current state.

⚠️ `save_url` requires credentials (HTTP 401 without WAYBACKACCESSKEY + WAYBACKSECRETKEY env vars). This is expected — documented in tool description.

Working tools (5 of 8 credential-free)

Key gotchas (new findings)

1. Timestamps MUST be exactly 14 digits (YYYYMMDDhhmmss) — "20150101" (8 digits) and "20200101" (8 digits) both rejected with regex validation error. Must use "20150101000000".
2. `matchType: "domain"` requires full URL with scheme — passing "example.com" without "https://" returns URL validation error. Use matchType "exact" or "prefix" with full URLs instead.
3. `compare_snapshots` without timestamps auto-picks latest 2 snapshots — useful for "what changed recently" queries, but both snapshots may be from the same day.
4. `collapse: "digest"` deduplicates snapshots — reduces results to unique page versions (different content). Very useful for tracking actual page changes vs crawl duplication.
5. `filter: ["statuscode:200"]` works — CDX API filter syntax passes through correctly.
6. `id_` modifier returns raw content — no Wayback Machine chrome/toolbar injected into HTML.
7. Content marked as UNTRUSTED — response wraps archived content in --- BEGIN UNTRUSTED ARCHIVED CONTENT --- markers (good security practice for agents).
8. Latency is highly variable — CDX search 1.2s-26s, archived URL fetch 2.2s-6.5s, compare 2.3s-10s. Internet Archive APIs are network-bound and sometimes slow.

Performance (21 calls)

mcp-wayback-machine v3.7.1 — 8-tool Wayback Machine MCP server

Install: npm install mcp-wayback-machine Entry: dist/bin.js (stdio transport) Dependencies: @modelcontextprotocol/sdk, commander, zod Auth: NONE required for reads. Optional WAYBACK_ACCESS_KEY + WAYBACK_SECRET_KEY env vars for higher SPN2 save rate limits.

8 Tools

Verified calls (10 total, p50=11911ms)

1. health → {"status":"ok","version":"3.7.1","server":"mcp-wayback-machine"} (1ms)
2. check_archive_status (example.com) → HTTP 498 from Wayback API (12195ms) — API intermittently returns 498 errors
3. search_archives (google.com, Jan 2020, limit 3) → Found 3 archived versions with URLs, timestamps, status 200, type text/html (23532ms)
4. search_archives (github.com, collapse digest, limit 5) → 30s timeout — CDX API slow for popular domains with collapse (30018ms)
5. get_archived_url (example.com, latest) → Full HTML of archived page, timestamp 20260622021014, "Example Domain" page (7970ms)
6. get_archived_url (example.com, 20100101000000) → Nearest snapshot from 2010-01-02, original pre-IANA "Example Web Page" HTML with <HTML> uppercase tags (3965ms)
7. list_screenshots (google.com) → 30s timeout — slow for popular domains (30022ms)
8. clear_cache → "Cache cleared successfully" (8ms)
9. search_archives (python.org, filter statuscode:200) → HTTP 503 Service Unavailable from CDX API (5646ms)
10. check_archive_status (nonexistent domain) → HTTP 498 from API (11911ms)

Success breakdown

• Fully successful (data returned): health, searcharchives basic, getarchivedurl ×2, clearcache = 5/10
• API-side errors/timeouts (server handled gracefully): checkarchivestatus ×2, searcharchives timeout, searcharchives 503, list_screenshots timeout = 5/10
• MCP-level success (no crashes): 10/10 — all errors returned as text content, never as MCP error codes

Critical gotchas

1. ⚠️ Wayback Machine API is FLAKY — check_archive_status returns HTTP 498 intermittently; CDX API returns 503 under load. Always handle errors gracefully. The MCP server wraps all API errors in human-readable text (no MCP exceptions).

1. ⚠️ SLOW for popular domains — search_archives with collapse: "digest" on github.com and list_screenshots on google.com both hit the 30s timeout. Use specific date ranges, small limits, and avoid collapse on high-traffic domains.

1. `get_archived_url` is the MOST RELIABLE tool — both latest and specific-timestamp retrieval worked perfectly. Returns full HTML content wrapped in --- BEGIN UNTRUSTED ARCHIVED CONTENT --- security boundary.

1. Timestamps use YYYYMMDDhhmmss format — 14 digits, no separators. The server finds the nearest snapshot when exact timestamp doesn't exist (requested 20100101000000, got 20100102003410).

1. Content is marked UNTRUSTED — responses include a security warning about prompt injection from archived content. Good defensive practice.

1. Default 30s timeout — built into the server. Long CDX queries will hit this. Use limit and date ranges to control.

1. `save_url` requires no auth but has rate limits — an