Use it from an LLM¶

naina ships an MCP server, so an assistant can read documents directly instead of being handed a wall of pre-extracted text.

Setup¶

cd mcp && npm install && npm run build

Then register it. For Claude Code:

claude mcp add naina -- node /absolute/path/to/naina/mcp/dist/index.js

Or by hand, in your MCP client's config:

{
  "mcpServers": {
    "naina": {
      "command": "node",
      "args": ["/absolute/path/to/naina/mcp/dist/index.js"]
    }
  }
}

Tools¶

`read_document`¶

Returns markdown. This is the one to reach for by default — it is what you want in a context window.

Argument	Type	Notes
`path`	string	absolute path to an image
`tier`	string	`tiny` / `small` / `medium`, default `small`

`read_document_detailed`¶

Returns JSON: every line with its confidence and its quad, plus layout regions and reading order.

Use this when the model needs to reason about where something is — checking a signature block, pulling a figure caption, or deciding whether a low-confidence line is trustworthy.

Why markdown by default¶

A bag of text loses the difference between a heading and a paragraph, and between a table cell and prose. Markdown keeps that structure in a form models already read well, at almost no token cost.

Protocol version¶

The server speaks MCP 2025-11-25.

The newer 2026-07-28 revision suits naina well — it is stateless request/response, and this server holds no session state — but @modelcontextprotocol/sdk@1.30.0 (the newest published) tops out at 2025-11-25, and a client asking for the newer revision negotiates down. Measured, not assumed. Moving up should be a dependency bump rather than a rewrite; tracked in the roadmap.