MCP server: One MCP server.

TGGET Niche Research

@tggetMCP server · External

Validate app and SaaS ideas: search demand, competitors, app store data and a verdict.

Freeno charge per call

One address for any MCP client. No sign-in needed on this server.

A niche report: verdict and the factors behind it
From the README

What it does

6 read3 act1 other

Grouped by tool name.

What it costs

Free

Odel charges nothing per call. The publisher may charge for its own service.

Tools

10 · grouped by tool name · checked 11 Oct 2026

Read · 6Look things up

  • get_research
  • get_report
  • list_research
  • compare_research
  • list_projects
  • get_project

Act · 3Change something

  • start_research
  • run_signal
  • update_project

Other · 1Neither hinted nor named

  • follow_niche
Try it in the Inspector
README5 min read

TGGET Niche Research MCP server

Remote Model Context Protocol server of TGGET, a niche research tool for mobile apps and online services. It lets an AI assistant check an idea against measured data instead of guessing from a description.

Give it a search phrase or a description of an idea in your own words. It returns monthly search demand, sub-niches, competitor sites from search results, App Store apps with what their users complain about, community signals and a final report with a verdict: worth building, needs validation or no niche. The verdict is computed by rules from five measured factors, so the same numbers always give the same answer; the language model only explains it.

  • Endpoint: https://tgget.io/mcp (Streamable HTTP)
  • Authentication: personal access token, sent as a Bearer token. The server tells what it is and lists its tools without a token; every call needs one
  • Name in the MCP Registry: io.tgget/niche-research
  • Documentation: tgget.io/docs/api

A niche report: verdict and the factors behind it

What you can ask

  • "Validate this idea: an app that reminds you to water plants and explains how to care for them."
  • "Is there demand for an expense tracker for couples? Who are the competitors?"
  • "Research the niche 'invoice generator' and give me the report as Markdown."
  • "Compare my last three researches and tell me which idea to start with."
  • "Follow the niche of my research about habit trackers and show what changed since last month."

Getting a token

  1. Create an account at tgget.io or in the Android app and confirm your email. A free account is enough.
  2. Open Settings, section "API and AI assistants", and create a token. It is shown once.

Tokens expire after a year and are revoked when you change your password or close your sessions. An account holds up to five tokens.

Connecting

Replace <token> with your token. Ready-made files are in examples.

Claude Code

bash
claude mcp add --transport http tgget https://tgget.io/mcp --header "Authorization: Bearer <token>"

Cursor

~/.cursor/mcp.json or .cursor/mcp.json in a project:

json
{
  "mcpServers": {
    "tgget": {
      "url": "https://tgget.io/mcp",
      "headers": { "Authorization": "Bearer <token>" }
    }
  }
}

VS Code

.vscode/mcp.json:

json
{
  "inputs": [
    { "type": "promptString", "id": "tgget-token", "description": "TGGET access token", "password": true }
  ],
  "servers": {
    "tgget": {
      "type": "http",
      "url": "https://tgget.io/mcp",
      "headers": { "Authorization": "Bearer ${input:tgget-token}" }
    }
  }
}

Codex

~/.codex/config.toml, with the token in the TGGET_TOKEN environment variable:

toml
[mcp_servers.tgget]
url = "https://tgget.io/mcp"
bearer_token_env_var = "TGGET_TOKEN"

Clients that only run local servers

Use the mcp-remote bridge:

json
{
  "mcpServers": {
    "tgget": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://tgget.io/mcp", "--header", "Authorization: Bearer <token>"]
    }
  }
}

Tools

ToolWhat it doesParameters
start_researchStarts a research from a search phrase or a description of an idea. A phrase is scanned directly; a description is first turned into real search queries and the strongest one is scanned.text (required), deep, locale
get_researchCurrent state and findings: status, headline numbers, ranked candidate phrases, competitor overview, child researches, skipped stages.id (required)
get_reportThe final niche report: verdict, customer pains, MVP, napkin economics, SEO and ASO plans, name options, landing page draft, distribution channels, validation steps and risks.id (required), format (markdown or json)
list_researchThe most recent researches of the account, newest first.limit (1 to 50)
run_signalAdds one check to a finished research: competitors for a phrase, demand history, App Store, reviews, community signals, paid search metrics, or a rebuilt report.id, kind (required), phrase, locale
compare_researchPuts two to five finished researches side by side and marks the better value of every measure. The marks come from the numbers, not from the language model.ids (required)
follow_nicheStarts following the niche of a finished research that has a report: it is scanned anew every month and a digest of what changed is sent to the account. Monthly checks do not spend researches.research_id (required)
list_projectsThe niches the account follows, with the latest numbers and the date of the next check.limit (1 to 50)
get_projectA followed niche with the history of its checks and what changed since the previous one.id (required)
update_projectPauses a followed niche, follows it again or stops following it.id, status (active, paused or deleted) (required)

locale sets the language of the conclusions and the report; en is the one available now. The market is chosen by the language of the phrase: search demand is measured in the country of that language. Texts meant for a site or an app store are always written in the language of the market.

Resources and prompt

NameWhat it gives
tgget://quotaResearches remaining on the plan and when the next one becomes available.
tgget://sourcesResearch kinds available right now and what each one costs.
find_niche (prompt)Walks the assistant through validating an idea: start, wait, read the evidence, write a verdict. Argument: idea.

How a research runs

Research runs in the background and takes a few minutes.

  1. Call start_research. It returns the research id.
  2. Poll get_research every 20 to 30 seconds until status is completed and pending is 0. Findings appear as they arrive.
  3. When the research lists a report, call get_report.

Limits and privacy

  • A research is counted when it starts. Checks added to an existing research with run_signal are free.
  • The free plan includes three researches to start and one more every 30 days. Paid plans are on the plans page and are bought in the mobile app; the same account works everywhere.
  • A free account follows one niche, Starter two, Pro ten.
  • Researches made on the free plan are public: the result may be listed in the open niche catalog. The catalog shows the researched phrase and its findings, never the description you typed. Researches on paid plans stay private.
  • The API and the MCP server share a limit of 60 requests per minute.
  • Search interest is evidence, not a customer count. Treat the verdict as a starting point for your own validation.

This repository holds the documentation and connection examples of the hosted server. The server itself runs at tgget.io.

Similar modules

Browse Market research