Creator Search Insights API: what exists in 2026

There is no public Creator Search Insights API. TikTok ships the tool in the app and at tiktok.com/csi, and has never documented an endpoint for it. The Research API does not return search topics, the Business API does not either, and there is no scope, partner tier or application that unlocks one.

What does exist is a way to call the same topic records from code: an MCP endpoint with a bearer key, one tool call per question. Below is the runnable Python, the JSON it returns, a field reference for every value in a topic record, and the follow-up calls that expand one topic into a cluster.

Why there is no Creator Search Insights API

Creator Search Insights is a creator-facing product, not a developer one. TikTok's developer portal is organised around four surfaces, and none of them touches search demand.

The Research API comes closest and still cannot answer the question. Its video query specification filters by keyword, hashtag, region and date, and returns videos. It has no concept of a search topic, a 7-day search count or a content gap. You can count how many videos mention sourdough. You cannot learn that 4.7 million people searched for it last week.

The broader map of every TikTok API and what each one is good for is in the TikTok API guide.

The endpoint that does return topics

TokConnect exposes the Creator Search Insights topic records, plus ordinary TikTok search, over the Model Context Protocol. It is one HTTP endpoint speaking Streamable HTTP and JSON-RPC.

MCP is not REST. There is no /search_topics URL to curl. You open a session, initialize, then call named tools with tools/call. Any MCP client library handles that sequence for you, which is the point of using one. If you want the same thing in a chat window rather than a script, the connect pages cover Claude, Codex, Cursor and Gemini CLI, and the TikTok MCP guide explains the protocol.

A Python example that runs

This uses the official mcp package. Install it with pip install mcp, put your key in TOKCONNECT_KEY, and run it.

import asyncio
import json
import os

from mcp import ClientSession
from mcp.client.streamable_http import streamablehttp_client

URL = "https://mcp.tokconnect.com/mcp"
HEADERS = {"Authorization": f"Bearer {os.environ['TOKCONNECT_KEY']}"}


async def main():
    async with streamablehttp_client(URL, headers=HEADERS) as (read, write, _):
        async with ClientSession(read, write) as session:
            await session.initialize()

            tools = await session.list_tools()
            print([t.name for t in tools.tools])

            result = await session.call_tool(
                "search_topics",
                {"keyword": "sourdough starter", "limit": 5, "language": "en"},
            )
            data = json.loads(result.content[0].text)

            for topic in data["topics"]:
                print(
                    f"{topic['searchVolume']:>10,}  "
                    f"{topic['videoCount']:>6}  "
                    f"{topic['queryText']}"
                )
            print("next offset:", data["nextOffset"])


asyncio.run(main())

import httpx2
from mcp.client.streamable_http import streamable_http_client

async with streamable_http_client(URL, http_client=httpx2.AsyncClient(headers={"Authorization": f"Bearer {KEY}"})) as (read, write):
    async with ClientSession(read, write) as session:
        ...

Two things to note. Tool results arrive as text content, so you parse result.content[0].text as JSON rather than reading a typed object. And list_tools() is worth calling once in development: the tool set grows, and the schemas it returns are the authoritative argument list.

Running it on 19 September 2026 printed this:

 2,923,707       0  dry sourdough starter and dough
 3,721,671       0  sourdough starter making and uses
 4,418,725       0  sourdough eating and starter care
 4,685,193       0  sourdough baking and starter care
 3,205,561     546  sourdough starter making and feeding
next offset: 5

Data: TikTok Creator Search Insights via TokConnect, pulled 19 September 2026. Open the snapshot.

The JSON search_topics returns

Here is one full record from that call plus the envelope around it, trimmed to a single topic:

{
  "topics": [
    {
      "queryId": "7668817079025287188",
      "queryText": "sourdough baking and starter care",
      "popularity": 87,
      "searchVolume": 4685193,
      "videoCount": 0,
      "trend7d": [9597, 106468, 291212, 442767, 1091572, 1611616, 4685193],
      "trendScore7d": [45, 70, 74, 78, 85, 85, 87],
      "categoryL1": "Gourmet",
      "categoryL2": "Food Tutorials",
      "isFavorited": false,
      "channel": "search",
      "tab": "SEARCH"
    }
  ],
  "hasMore": true,
  "nextOffset": 5,
  "count": 5,
  "channel": "search",
  "tab": "SEARCH"
}

That record is a content gap in JSON form: 4.69 million searches in a week against zero matched videos. The content gap guide explains what TikTok counts as a match.

Field reference

Keep an absent field absent when you store these. A videoCount of zero and a videoCount that never arrived are different states, and collapsing them ruins any content-gap filter you build on top.

Follow-up tools

Each of these takes the queryId from a topic record and costs one credit.

The useful pattern is one broad search_topics call, then follow-ups on the two or three topics that survive your filter. Calling audience_demographics on all ten results is nine wasted credits, because you will drop most of them on volume alone.

Pagination, credits and rate

search_topics takes limit from 1 to 20 and an offset. The response returns hasMore and nextOffset; pass that nextOffset straight back as offset for the next page. In the example above the response came back with nextOffset: 5, so the next page starts at 5. Do not compute the offset yourself from page numbers, because the server decides the step.

Billing is one credit per tool call, whatever the call returns, including a call that returns an empty list. Every account starts with 100 free credits, which is enough to build and test a real pipeline rather than one demo request.

Two practical notes. Keep the key out of source control and read it from the environment, as the example does. And if a call comes back with an upstream error, wait and retry once before failing the job: these are live TikTok reads, and a transient failure on one topic is not a reason to abandon the batch.

If you would rather ask questions than write a client, the desktop guide shows the same data through Claude or ChatGPT with no code at all.