MCP

The agent tools are an MCP server, for any agent or app that speaks MCP: Claude Code, Cursor, Claude Desktop and others.

With OpenAI, use the agent tools instead.

The OpenAI Responses API gives the model each MCP tool's text only, not the pictures of pages that view_pages and zoom_page return, so the agent cannot look at scans, charts or maps. Run the tools in your code with the agent tools, with the OpenAI Agents SDK or function calling, and the model sees the pages. Claude Code and Claude Desktop pass the pictures on.

Connect

Connect it once, at https://api.search.pageindex.ai/mcp with your API key. The agent sees all your documents, and finds the collection or document a question names. Searches are priced as any other; listing collections and documents and reading pages are free.

claude mcp add --transport http page-search \
  https://api.search.pageindex.ai/mcp \
  --header "Authorization: Bearer pis_live_..."

Choose the documents

Name a collection or document in the question to start there. To keep an agent to some documents whatever it is asked, such as one that answers your customers from one collection, add ?collection=col_..., or ?filter= with a metadata filter as JSON, to the address. See Choose the documents.

On your computer

Claude Desktop runs the server on your computer: pageindex-search-mcp comes with the Python SDK (pip install "pageindex-search[mcp]"), reads PAGEINDEX_SEARCH_API_KEY, and takes --collection or --filter to keep the agent to some documents.

{
  "mcpServers": {
    "page-search": {
      "command": "uvx",
      "args": ["--from", "pageindex-search[mcp]", "pageindex-search-mcp"],
      "env": { "PAGEINDEX_SEARCH_API_KEY": "pis_live_..." }
    }
  }
}

In Python, client.tools().mcp() gives the server, to run over stdio or serve over HTTP yourself.

Try it

The console's Agent playground asks an agent about your documents with these tools, showing each step and what it cost. Or ask about one of the examples, no sign-in needed.