TikTok MCP: research TikTok from your AI agent

A TikTok MCP server gives an AI agent tools it can call to retrieve TikTok data. With TokConnect, that includes search topics, videos, public creator profiles, comments, and related research. You ask the question in your existing agent; it chooses the tools and works with the responses.

This guide covers organic content research. For advertising operations, TikTok offers a separate TikTok for Business MCP server. TokConnect is an independent service.

What can your agent research?

The MCP supplies records. The agent can sort, calculate, compare and draft from those records. Someone still needs to watch a linked video to assess its visuals: TokConnect does not return a transcript or analyze the footage.

Videos

Search trends

Comments

A real topic search: walking pads

On 18 September 2026, we called keyword_research for walking pad, with three related-topic expansions and video retrieval off. It returned 20 matching topics. Here are the first three saved records.

For a product-demonstration brief, “products and usage” is the stronger candidate to review first; “stepper reviews” suggests a comparison format. That is an editorial recommendation from the topic wording, not a completed production decision. The lower listed-video count alone does not prove lower competition.

The response names searchVolume, popularity and videoCount. It does not establish a monthly period or geographic scope for the estimate. Our metrics guide explains those fields and shows a worked comparison.

Follow a question across tools

Research usually needs more than one lookup. Select a topic, retrieve relevant videos, choose a video, and retrieve its comments. Returned IDs connect those steps, so your agent does not have to guess which record you mean.

In a separate push-up-app example, a call requested 50 comments and returned 49. The completed analysis groups every returned comment and links the coding sheet. That is the difference between asking an agent for general marketing advice and giving it source material from the audience you are investigating.

Research TikTok videos about [product]. Find five relevant examples, list their captions and public counts, and retrieve up to 30 comments from two selected videos. Group repeated product questions with the source links. Suggest two questions our next demo should answer.

Set a call budget in the prompt if the task could expand. Every MCP tool call costs one credit. A composite research tool can make several upstream requests within that one call; requesting another tool or page adds another credit.

Connect your agent

1. Sign in to TokConnect and copy your API key. New accounts receive 100 credits.
2. Add https://mcp.tokconnect.com/mcp using your client's remote MCP settings and bearer authentication.
3. Confirm that the client lists TokConnect's tools, then run one small search before starting a larger workflow.

Choose the instructions for Claude Code, Codex, OpenClaw, Hermes, Cursor, or Gemini CLI. Claude Desktop has different authentication requirements; check its connection options first. Each guide states what was actually tested.

When to use MCP, an API, or TikTok itself

Use TikTok's own Creator Search Insights interface for an occasional manual topic check. Use an MCP connection when you want an agent to combine topic, video, profile and comment lookups in your existing workflow. Developers can also call the server from an MCP SDK; see the programmatic access guide.

Publishing videos, operating ads, and approved academic data access use other products. The TikTok MCP and API comparison explains which access path matches each job.

Complete an organic research run

Start with a question that can produce a decision. “Find ideas for our brand” sends an agent into an open-ended search. “Choose one of three product-demo angles for people who work from home” gives it a boundary. Name the product, intended viewer, decision, and the maximum number of topic, video, and comment calls before you connect.

The saved walking-pad retrieval supplies a worked sequence. It started with the phrase walking pad and retained three topic records. For a product demonstration, use “walking pad products and usage” as the candidate to review first. “Walking pad and stepper reviews” points toward a comparison format. The result supports that next research choice; it does not record a completed production decision.

1. Ask search_topics for the seed and retain the exact language, offset, retrieval date, topic label, and queryId.
2. Choose one topic against a written relevance rule. In this example, the rule is “shows how the product is used while working.”
3. Use related_videos or search_videos for a bounded set. Watch the selected source URLs before describing an opening, claim, or visual treatment.
4. Use video_comments only on a selected video and record its cursor and returned count. Code the returned rows before turning a repeated question into a test.
5. Write a short memo that separates the returned facts from the creative choice.

A useful final memo can read: “Use a product-in-use demo for the walking-pad angle. The selected topic label is ‘walking pad products and usage’; the choice came from a 18 September 2026 topic result. Before scripting, inspect five returned videos and one comment page. We have not measured conversion, buyer intent, or the full result set.” A writer can act on that memo and a reviewer can trace its limits.

Choose the client setup that you can finish

Use a remote-MCP client when the agent already supports remote servers and bearer authentication. Add the endpoint, enter the key in the client’s secret or header field, and confirm tool discovery before asking for research. The first successful result should be a small call such as search_suggestions for one phrase or list_languages. A topic report is a poor first test because it combines several upstream requests and makes a setup failure harder to isolate.

Use an MCP SDK only when you need code to control the research loop, save records to your own system, or apply your own review rules. Your program still performs the same protocol work: open the authenticated transport, initialize, list tools, call a named tool, then retain the response. The program gains control over retries and storage; it does not gain a different TikTok permission set.

Keep the credential in the client’s secure setting or a process environment variable. A browser-accessible page, a repository config file, and a shared prompt are poor places for it. If discovery works but a research call fails, save the tool name, non-secret arguments, timestamp, and returned error. Those four fields let support distinguish a connection problem from an upstream response.

Interpret results and failures

The local server also treats an empty video-search response carrying a search refusal signal as an error, not evidence that no videos match. Retry according to the service error, reduce parallel work, or stop and preserve the error in the log. Do not translate a transport or limit failure into a market conclusion.