A PDF MCP server for AI agents.

An MCP server gives an agent tools it can call without writing code. nupdf's is the same binary as its command line: every command is one tool, and every result is the command's JSON report.

Add it

$ cargo install --locked nupdf-cli
$ claude mcp add nupdf -- nupdf mcp

For a whole project, check in a .mcp.json. Other MCP clients take the same command and arguments.

{
  "mcpServers": {
    "nupdf": { "command": "nupdf", "args": ["mcp"] }
  }
}

What the agent gets

• One tool per command. info, extract, search, tables, render, forms, merge, split, redact, verify, diff and the rest, each taking one file or many.
• Structured results. Each call returns the report as text and as structuredContent. A failure is a tool error carrying nupdf's error report, with a stable code.
• An address for every span of text. Search returns ids like p3.b2.l4.s1:5-12, and other tools accept them.
• Budgets. max_chars stops at a whole block and returns a cursor to continue from.
• Limits. Every call is bounded: 300 seconds and 8 MB of report by default. A cancelled call stops.
• It keeps serving. A panic on one file is caught and reported; the server does not die with it.

How it did in the agent eval

Thirty graded tasks, five runs each, one small-model agent a run, each allowed its own tool only. The three MCP servers in it are nupdf's, PDF Oxide's and pymupdf4llm-mcp.

Tokens and wall time are medians over passed runs. The two other servers cannot write a PDF, so they skip the tasks that need one. Tasks, graders and every run

Prefer a skill?

The same commands work as a Claude Code skill, where the agent runs the CLI from its shell.

$ mkdir -p ~/.claude/skills/nupdf
$ nupdf guide --skill > ~/.claude/skills/nupdf/SKILL.md

In the eval, agents running the CLI from a shell solved as many tasks as those using the server, and used more tokens: a median 14,578 a task against 8,988.