Built MCP-first for Claude, Cursor & every agent

Query millions of LLM prompts in plain English

Promptbase indexes the prompt web in real time — every prompt, every citation, every source — so your go-to-market team (or agent) can just ask.

Open the console

MCP session · geoindexer

live data

“How many cold email prompts cite reddit?”

Searching the citation graph…

Works everywhere agents live

Claude

Cursor

DSH

ChatGPT

Windsurf

5,354Prompts indexed

33,618Citations recorded

3Targets classified

179msQuery latency

What it does

Not another dashboard. A queryable citation graph.

One scrape serves many tenants: prompts are embedded once, harvested on a schedule, and every answer keeps the raw response so citations can be re-derived at any time.

Natural language queries

Describe your product in a sentence. Hybrid BM25 + vector search fuses both rankings with RRF — no filters UI, no SQL.

MCP-native from day one

The index ships as an MCP server: count_prompts_citing, search_prompts, corpus_stats, top_cited_domains. Claude or Cursor can call them directly.

Freshness you can see

Every answer records when it was fetched. "Cites X" means a recent run actually cited it — not a stale scrape from last quarter.

Citation graph, not mentions

Targets are classified at insert time: LinkedIn Pulse, Reddit, Medium, or nothing. Query by what the model cited, not by keyword.

Opportunity scoring

Adjacency × citation presence × search volume, with a length prior when volume is unknown — so short, answerable prompts float up.

API + export

Five JSON endpoints, Zod-validated, single API key. Export a ranked prompt set as a file for your content or GTM team.

How it works

From a product blurb to cited prompts

We harvest real AI answers

Prompt corpus embedded and indexed, then harvested through ChatGPT web with citations parsed and classified — raw responses kept for replay.

You ask in plain English

Through MCP, the HTTP API, or the web console. One query serves every seat — this is a shared index, not a per-brand dashboard.

Get ranked, cited prompts

Each result carries the prompt, the exact URLs cited, the engine, and when it was last seen.

MCP-first architecture

Give your agent a search engine for prompts, not a UI.

The index ships as an MCP server — the HTTP API and the console are thin clients on top. Any agent that speaks MCP can call count_prompts_citing, search_prompts, or top_cited_domains directly.

Six tools: citations, search, corpus stats, prompt detail, domains, pipeline
Runs against the same Postgres as the API — no second data path
Widened evidence window: any of the last 3 runs per prompt counts as a citation

Corpus health Try a query

mcp.config.json

{
  "mcpServers": {
    "geoindexer": {
      "command": "npx",
      "args": ["-y", "tsx", "scripts/mcp-server.ts"],
      "cwd": "/path/to/geo-indexer",
      "env": { "EMBED_PROVIDER": "openrouter" }
    }
  }
}

Stop guessing what gets asked about your category.

Describe your product, see the prompts that already cite the places you want to be cited.

Run a query View corpus health

Or join the waitlist