Tiktok Kit

A Claude Code kit that reads your TikTok ads and organic data over MCP, works out what is actually happening in the account, and produces a prioritized growth direction.

Read-only. The kit never writes to TikTok. It analyzes and recommends; you execute in Ads Manager.

Install

npm install -g tiktok-kit
tk init

tk init does the whole setup: installs the kit, asks your output language, registers TikTok's official MCP server in Claude Code, and opens your browser to authorize.

No API key, no developer app, no approval wait.

Then open Claude Code and run:

/tk:connect

If you skipped the browser step, or 30 days have passed and authorization expired:

tk auth

Commands

Command	Does
tk init [-g]	install, register the server, and authorize — the whole setup
tk init --layered	connect the lighter ~40-tool server instead of the full ~400
tk init --no-connect	install only, register nothing
tk init --no-auth	register the server but skip the browser step
`tk lang [vi	en]`
tk auth	authorize with TikTok — also how you re-auth every 30 days
tk auth --no-browser	print the URL instead of opening a browser (SSH)
tk connect	reconnect, or switch tool surface
tk connect --advanced	add other sources — organic data lives here
tk status	install mode, payload version, connected data planes, query budget
tk remove [-g]	remove assets and unregister

Scripted and CI installs skip both the language prompt and the browser step automatically, so nothing hangs.

Authorization lasts 30 days. When an ads plane that worked last month goes quiet, tk auth is almost always the fix.

Skills

/tk:connect → /tk:account → /tk:diagnose → /tk:creative → /tk:direction
                                       ↘  /tk:research  ↗

Skill	Does
/tk:connect	detect which data planes are reachable, guide setup
/tk:account	baseline the account, check tracking health, emit dataTrust
/tk:diagnose	find what is wrong across paid and organic, including cross-plane mismatches
/tk:creative	explain why videos won or lost, from the actual media
/tk:research	market research paired with your own account history
/tk:direction	at most five ranked moves, each with a kill criterion

/tk:research takes subcommands: trend, competitor <handle>, audience, format.

Agents

tiktok-data-analyst (descriptive only) · tiktok-creative-strategist · tiktok-trend-researcher · tiktok-growth-strategist (the only one permitted to recommend).

That separation is deliberate: advice carried inside a data report inherits the authority of the data.

Connecting Your Data

tk init connects TikTok's official MCP server. Nothing to configure: the entry is a URL, and authorization happens in the browser.

The official server covers ads only. For follower, video, and completion data you need a second source:

tk connect --advanced

That menu holds a hosted vendor (ads + organic, needs an API key, a third party holds your TikTok token), a self-hosted server (ads only, needs a TikTok developer app that TikTok must approve), and a content source for research. Each states its trade-off before you authorize anything.

Credentials

The official path stores none — .mcp.json holds only a URL.

Advanced sources that do need a key read it from a muted terminal and write it to .mcp.json, then add that file to .gitignore because it is plaintext. Setup lives in the CLI precisely so a key never lands in a Claude conversation transcript. .tk.json records connection state only, never a token.

You still need real history

The kit compares 28 days against the previous 28. An account younger than about two months produces noise, not insight — no amount of tooling fixes that.

What This Kit Does Not Do

• Write to TikTok — no campaign creation, status changes, budget edits, or uploads
• Publish or schedule content
• Connect to TikTok's official Agentic Hub — it exposes no public endpoint for external clients yet. When it does, only two files need to change

Honest Limitations

• Thresholds are provisional. Defaults were authored from documentation and have never been calibrated against a real account. /tk:direction shows that confidence rather than hiding it.
• Rules are instructions, not enforcement. Nothing blocks a violation at runtime. CI gates enforce the artifacts — that templates demand citations, that thresholds carry provenance, that vendor names stay confined — not the model's behavior.
• Query budget can be overrun. Metered vendors bill per query and a free tier is small. Set budget.maxQueriesPerRun in .tk.json to match your plan.

Development

npm run manifest   # regenerate release-manifest.json after payload changes
npm run check      # run all 9 quality gates

Contributor rules: docs/kit-contract.md.

License

MIT