Setup and reference

The server speaks stdio. Clone the repository, install, and point your client at the stdio script.

npm install
npm run mcp:stdio

For Claude Desktop, add an entry to your MCP config pointing at that command, with the repository as the working directory.

{
  "mcpServers": {
    "pepiros": {
      "command": "npm",
      "args": ["run", "mcp:stdio"],
      "cwd": "/path/to/pepiros"
    }
  }
}

There is no published npm package yet, so npx pepiros-mcp will not resolve. Remote HTTP with OAuth is designed but not built.

Every claim carries one of three states. They are not degrees of confidence in the research; they describe how well the claim matched the text it cites.

• quote located: the quote scored 0.92 or above against a real chunk of the source. Page and quote are both shown.
• paraphrase: scored between 0.75 and 0.92. Close, but not a verbatim match.
• unsupported: below 0.75. The anchor is dropped and the citation marker is stripped on re-verification.

A claim with no checked excerpt at all is badged inference. Nothing is ever badged “verified”. See how it works for the reasoning.

Anchors use stable, human-readable ids so a citation stays meaningful across a conversation: C7 for a text chunk, F3 for a figure, N12 for a numeric ledger row.

An anchor can span multiple rects on a page, which is how a quote that wraps across a column break still highlights correctly.

Eight tools are registered: list_papers, search_paper, verify_claim, get_outline, get_node, create_node, find_contradictions, and paper_facts. Full argument tables are on the MCP page.

create_node re-verifies server-side and never trusts a client-asserted tier. An agent cannot write a claim into the graph badged “quote located” by simply saying it is.