tiktok-public-research

TikTok public-research MCP (tiktok)

Read-only access to public TikTok data via the self-hosted tiktok MCP server (~/source/repos/tiktok-mcp, a FastMCP wrapper over davidteather/TikTok-Api 7.3.3). Same self-hosted ethos as the other social MCPs — but public-scrape only, no login.

Core principle

There is no authentication. The server logs into nothing and exposes nothing about your account. An optional ms_token (an anti-bot cookie copied from a logged-in tiktok.com session) and an optional residential proxy only make the scraper look less bot-like — they are not credentials and do not unlock private data. Everything this server reads is data any logged-out visitor could see.

Tools (all read-only, public data)

Tool	Reads	Notes
tiktok_trending(count)	trending feed	works keyless; geo-localized to the egress IP/proxy
tiktok_user_info(username)	a creator's public profile + stats	needs ms_token (keyless 403s)
tiktok_user_posts(username, count)	a creator's public posts	needs ms_token
tiktok_user_liked(username, count)	a creator's public likes	only if that user made likes public (usually empty)
tiktok_hashtag(tag, count)	recent videos for a hashtag	needs ms_token
tiktok_video_details(url_or_id)	one video's metadata	URL preferred over numeric id
tiktok_comments(url_or_id, count)	top-level comments on a video	needs ms_token
tiktok_related_videos(url_or_id, count)	videos TikTok relates to one video	needs ms_token
tiktok_search_users(query, count)	public user accounts by name	user search only — 7.x has no generic video search

count is clamped to [1, 50]. Errors come back structured ({error, detail, hint}), never as a crash.

Reliability — the real constraint (not code, bot-detection)

• tiktok_trending works with no setup. Every other tool typically returns EmptyResponseException until an ms_token is set — TikTok bot-walls user/hashtag/ video/comment endpoints for an unauthenticated scraper. This is expected, not a bug.
• TIKTOK_MS_TOKEN — copy the msToken cookie value from a logged-in tiktok.com browser session into the gitignored .env (or pass via claude mcp add -e). Refresh it if calls start returning empty. It is a nonce, not a login — but treat it as private (gitignored, never logged, never committed).
• TIKTOK_PROXY (a residential proxy URL) — required on a datacenter/cloud egress IP (TikTok blocks datacenter IPs almost instantly). Not needed on a home residential IP. Same rule and same Byteful sticky-RO endpoint used for IG/FB on cloud egress.
• TIKTOK_BROWSER — chromium (default), webkit, or firefox. A different engine sometimes evades detection; webkit alone did not fix the keyless user/hashtag block in testing, so reach for ms_token first.

Hard limits (state them, never imply otherwise)

• No own-account data: your feed, your saved/favorites, your following, your likes — none are reachable (there's no session). For your own posts/profile use the official Display API instead, not this scraper.
• No DMs — no unofficial library exposes the TikTok message inbox; do not imply DM search.
• No writes — no posting, liking, following, commenting. Read-only by construction.
• Unofficial-API risk: heavy scraping from one IP gets that IP blocked (EmptyResponseException); it does not ban a TikTok account (nothing is logged in). Keep usage human-paced.

Setup

Server: ~/source/repos/tiktok-mcp (Python 3.12 .venv, TikTokApi + mcp[cli], Playwright chromium/webkit installed). Registered user-scope as tiktok:

claude mcp add tiktok --scope user -e TIKTOK_MS_TOKEN=<msToken> -- ^
  C:\source\repos\tiktok-mcp\.venv\Scripts\python.exe C:\source\repos\tiktok-mcp\server.py

Tools load on the next Claude Code restart. Per the toolchain's MCP policy, consider adding tiktok to each project's disabledMcpServers and enabling it only where needed (same as the drug MCPs). Full build/setup detail: docs/social/TIKTOK-MCP.md.

Pairs with the broader social MCP reference: [[social-media-toolchain]].