MCP server

A stateless MCP server at https://mcp.arxivv.org/mcp (also https://arxivv.org/mcp) with seven tools for reading papers by anchor.

arXivv runs a Model Context Protocol server built on the document tree.

Property	Value
Endpoint	`https://mcp.arxivv.org/mcp`
Alias	`https://arxivv.org/mcp`, proxied to the endpoint above
Transport	Streamable HTTP, JSON responses
Sessions	Stateless: every call stands alone
Auth	None
Tools	All read-only
Rate limit	Requests through `arxivv.org/mcp` share the HTTP rate limit

Connect

{
  "mcpServers": {
    "arxivv": {
      "type": "http",
      "url": "https://mcp.arxivv.org/mcp"
    }
  }
}

With Claude Code:

claude mcp add --transport http arxivv https://mcp.arxivv.org/mcp

paper is any paper reference: 1706.03762, 1706.03762v7 or hep-th/9901001. Without a version, the latest is used.

Tool	Input	Output
`outline`	`{ paper, wait? }`	`Outline` as JSON
`read`	`{ paper, anchor, maxTokens? }`	One section’s Markdown, heading included
`cite`	`{ paper, anchor }`	`{ text, url, anchor }`: plain text of one paragraph or section, its section URL and resolved id
`references`	`{ paper, id? }`	`{ idv, references: Reference[] }`, or one `Reference` when `id` is given
`figure`	`{ paper, anchor }`	`FigureJson`, or `TableJson` for a table
`equation`	`{ paper, anchor }`	`{ id, number, latex }`; `number` is `null` when the equation is unnumbered
`search_paper`	`{ paper, query, limit? }`	`[{ anchor, sectionId, snippet }]`: the best-matching paragraphs

Tool	Argument	Accepts
`outline`	`wait`	Integer 0–60: seconds to wait for inferred fields, as `Prefer: wait=N`. See first request
`read`	`anchor`	Section id (`S3.SS2`), alias (`S3.2`), `A1` or a title slug (`introduction`)
`read`	`maxTokens`	Integer 100–100000, default 8000 (chars/4)
`cite`	`anchor`	Paragraph id (`S3.SS2.p3`) or section id, alias or slug
`references`	`id`	Reference id (`bib.bib12`) or printed number (`12` or `[12]`)
`figure`	`anchor`	Figure id (`S3.F1`) or `F<n>`; table id (`S6.T2`) or `T<n>`
`equation`	`anchor`	Equation id (`S3.E1`) or `E<n>`
`search_paper`	`query`	Words to look for, e.g. `label smoothing`
`search_paper`	`limit`	Integer 1–50, default 5

read truncation: output longer than maxTokens × 4 characters is cut and ends with …[truncated: N more tokens; fetch /1706.03762v7/sections/S3.2.md].
cite URL: url is https://arxivv.org/<idv>/sections/<sectionId>.md, where <idv> is the resolved version and <sectionId> the enclosing section’s id. anchor is the cited node’s id. For example, S3.SS2.p1 returns url: "https://arxivv.org/1706.03762v7/sections/S3.SS2.md" and anchor: "S3.SS2.p1". The Markdown has no in-page anchors, so the URL carries no fragment.
figure resolves tables too: T<n> and table ids return a TableJson with its cells.
search_paper ranking: the query is split on whitespace into lowercase terms. Paragraphs that contain more distinct terms rank first, then those with more total occurrences. Each snippet is about 150 characters either side of the first match. It searches one paper, by substring; it is not semantic.
Errors: a failed call returns isError: true with text such as not-found: Unknown paper 1706.99999 (the HTTP error code and detail).

outline({ paper: "1706.03762" }): title, TL;DR, contributions and the section tree with token counts.
read({ paper: "1706.03762", anchor: "S3.2" }): only the section you need.
search_paper({ paper: "1706.03762", query: "label smoothing" }): find paragraph anchors.
cite({ paper: "1706.03762", anchor: "S3.SS2.p1" }): exact text plus a stable URL to quote.
figure({ paper: "1706.03762", anchor: "F1" }) or equation({ paper: "1706.03762", anchor: "E1" }) as needed.

A typical question costs ~3K tokens this way, against ~15K for the full paper.

{
  "method": "tools/call",
  "params": {
    "name": "equation",
    "arguments": { "paper": "1706.03762", "anchor": "E1" }
  }
}

{
  "id": "S3.E1",
  "number": "1",
  "latex": "\\mathrm{Attention}(Q,K,V)=\\mathrm{softmax}\\left(\\frac{QK^{T}}{\\sqrt{d_{k}}}\\right)V"
}

Tool outputs follow schema 1.