arxiv-mcp-server

Public Hosted Server: https://arxiv.caseyjhand.com/mcp

Tools

Four tools for searching and reading arXiv papers:

arxiv_search

Search for papers using free-text queries with field prefixes and boolean operators.

• Field prefixes: ti: (title), au: (author), abs: (abstract), cat: (category), all: (all fields)

• Boolean operators: AND, OR, ANDNOT

• Optional category filter, sorting (relevance, submitted, updated), and pagination

• Category accepts a leaf code (cs.CL) or a whole archive (astro-ph, cs, math) — a bare archive covers its subject classes plus the legacy flat papers filed before it was subdivided

• submitted_from / submitted_to bound the submission date (inclusive, UTC YYYY-MM-DD). Consecutive windows cover the matches with no gap — a paper submitted exactly at a midnight seam falls in both, so de-duplicate by ID — which is how to reach results past the 10,000 pagination ceiling

• Echoes back the query as actually searched, with every filter folded in — replaying it reproduces the same result set

• Returns up to 50 results per request with full metadata including abstract

arxiv_get_metadata

Fetch full metadata for one or more papers by known arXiv ID.

• Batch fetch up to 10 papers in a single request

• Accepts both versioned (2401.12345v2) and unversioned (2401.12345) IDs

• Legacy ID format supported (hep-th/9901001)

• Reports not-found IDs separately from found papers

arxiv_read_paper

Read the full body of an arXiv paper.

• Tries native arXiv HTML first, then ar5iv, then text extracted from the PDF — the source field reports which one answered

• Strips HTML head/boilerplate and collapses MathML to dollar-delimited LaTeX ($…$ inline, $$…$$ block) so the character budget targets paper content

• Returns raw HTML — no parsing or extraction; the LLM interprets content directly. PDF-extracted bodies are plain text: prose is reliable, but math, tables, and heading structure flatten

• max_characters defaults to 100,000; pass null for the whole paper in one call. Raw HTML can be 500KB-3MB+ for math-heavy papers, which is more than most clients accept in a single tool result — page with start instead

arxiv_list_categories

List arXiv category codes and names for discovery.

• ~155 categories across 8 top-level groups (cs, math, physics, q-bio, q-fin, stat, eess, econ)

• Optional group filter to narrow results

• Static data — always succeeds

Related MCP server: Research Server

Resources

Features

Built on @cyanheads/mcp-ts-core:

• Declarative tool definitions — single file per tool, framework handles registration and validation

• Unified error handling across all tools

• Pluggable auth (none, jwt, oauth)

• Structured logging with optional OpenTelemetry tracing

• Runs locally (stdio/HTTP) from the same codebase

arXiv-specific:

• Read-only, no authentication required — arXiv API is free, metadata is CC0

• Rate-limited request queue enforcing arXiv's 3-second crawl delay

• Adaptive cooldown on rate-limit (5s → 10s → 20s → 30s), honors Retry-After

• Retry with exponential backoff for transient failures

• Content fallback chain: native arXiv HTML → ar5iv → PDF text extraction (both HTML renders run LaTeXML, so they tend to fail together; the PDF is the artifact every paper has, and it also covers an ar5iv outage rather than letting one fail the read)

• Full arXiv category taxonomy embedded as static data

• Optional local OAI-PMH metadata mirror (SQLite + FTS5) — opt-in, eliminates rate-limit exposure for arxiv_search and arxiv_get_metadata. See Optional: Local Mirror.

Getting Started

Public Hosted Instance

A public instance is available at https://arxiv.caseyjhand.com/mcp — no installation required. Point any MCP client at it via Streamable HTTP:

{
  "mcpServers": {
    "arxiv-mcp-server": {
      "type": "streamable-http",
      "url": "https://arxiv.caseyjhand.com/mcp"
    }
  }
}

Self-Hosted / Local

Add to your MCP client config (e.g., claude_desktop_config.json):

{
  "mcpServers": {
    "arxiv-mcp-server": {
      "type": "stdio",
      "command": "bunx",
      "args": ["@cyanheads/arxiv-mcp-server@latest"]
    }
  }
}

Prerequisites

• Bun v1.3.0 or higher.

Installation

1. Clone the repository:

git clone https://github.com/cyanheads/arxiv-mcp-server.git

2. Navigate into the directory:

cd arxiv-mcp-server

3. Install dependencies:

bun install

Configuration

All configuration is optional — the server works out of the box with sensible defaults.

Running the Server

Local Development

• Build and run:

  bun run build
  bun run start:http   # or start:stdio

• Run checks and tests:

  bun run devcheck     # Lint, format, typecheck, audit
  bun run test         # Vitest

Optional: Local Mirror

For self-hosted deployments behind a single egress IP, arXiv's ~3-second per-IP crawl delay serializes concurrent users. An optional local mirror eliminates rate-limit exposure for arxiv_search and arxiv_get_metadata by serving from a SQLite + FTS5 store harvested via OAI-PMH. arxiv_read_paper continues to use the live API — full-content harvest is forbidden by arXiv's data policy.

Disabled by default. To enable:

# 1. Cold-start harvest (~4.4h sequential, resumable from checkpoint). One-time per installation.
bun run mirror:init

# 2. Enable the mirror.
export ARXIV_MIRROR_ENABLED=true

# 3. Start the server — reads switch to the mirror once the harvest completes.
bun run start:http

Daily incremental refresh (small delta; duration depends on arXiv's OAI-PMH page pacing) via:

bun run mirror:refresh   # wire to cron / systemd timer / launchd, OR
                         # set ARXIV_MIRROR_REFRESH_CRON to schedule it in HTTP mode (spawned as a child process)
bun run mirror:verify    # schema version + PRAGMA integrity_check / quick_check

Schema upgrades. The mirror records a schema version and migrates itself in place the first time a newer server opens it — never a re-harvest, and never a separate operator step. The upgrade that added comment and journal_ref to the full-text index (#37) rebuilds that index from the rows already stored, so co: and jr: searches resolve against a mirror harvested before it. The rebuild runs at startup, before the store answers its first read, and logs mirror migration v2→v3 (fts rebuild) progress lines throughout — on a full-corpus mirror, expect the first start after the upgrade to take noticeably longer than usual. An interrupted rebuild is repeated on the next open rather than left half-applied. bun run mirror:verify prints the schema version the file carries and exits non-zero if a migration never completed.

Behavior notes. Ranking divergence: FTS5 BM25 differs from arXiv's internal ranking, so sortBy=relevance against the mirror returns a different top-K than the live API. Queries sorted by submitted descending within ARXIV_MIRROR_RECENT_DAYS_LIVE days route to the live API to cover the nightly-update gap. Refresh resilience: after the initial cold harvest completes, an in-progress or failed daily refresh keeps serving the existing dataset from the mirror — arxiv_search and arxiv_get_metadata don't drop to the live API during the refresh window (#21). The scheduled HTTP-mode refresh runs in a child process, so the harvest's synchronous SQLite writes never block the request event loop — search and metadata stay responsive throughout (#22). The mirror stores the latest version only; per-version reads continue to use the live API. See #12 for the full design.

Docker

docker build -t arxiv-mcp-server .
docker run -p 3010:3010 arxiv-mcp-server

Project Structure

Development Guide

See CLAUDE.md for development guidelines and architectural rules. The short version:

• Handlers throw, framework catches — no try/catch in tool logic

• Use ctx.log for domain-specific logging

• Rate limiting is managed by ArxivService — don't add per-tool delays

• arXiv API returns HTTP 200 for everything — check content-type and response body

Contributing

Issues and pull requests are welcome. Run checks before submitting:

bun run devcheck
bun test

License

Apache-2.0 — see LICENSE for details.