asta-skill — Semantic Scholar via Ai2 Asta MCP

English · 中文 · Asta MCP Overview · Request API Key

Ai2’s Asta MCP server (Semantic Scholar) runs on any MCP-capable agent — this repo adds an optional pure instruction-pack skill on top that turns natural-language research questions into well-formed tool calls: it routes intent → tool, fills safe defaults, chains workflows, and warns you off the usual pitfalls. Register the MCP server and you can call Asta from Claude Code, Cursor, Codex, Windsurf, Trae, Qoder, CodeBuddy/WorkBuddy, Hermes, opencode, OpenClaw, pi-mono, and anything else that speaks MCP; add the skill (on any Agent Skills-compatible host) for sharper results.

What it does

• Search the Semantic Scholar academic corpus by keyword, title, author, or full-text snippet
• Look up a paper from any ID (DOI, arXiv, PMID, PMCID, CorpusId, MAG, ACL, SHA, URL)
• Traverse citations — find who cited a given paper, with filtering and pagination
• Batch-lookup multiple papers in one call via get_paper_batch
• Snippet search — retrieve ~500-word passages from paper bodies for evidence grounding
• Author discovery — find researchers and list their publications
• Zero-code integration — the skill is a pure instruction pack; all I/O goes through the Asta MCP server
• Triggers automatically whenever the user asks for papers, citations, academic search, or literature discovery and Asta tools are registered

Multi-Platform Support

Anything that speaks MCP can call Asta — the skill is just an optional Agent Skills layer on top. Verified on Claude Code, Codex, Cursor, Devin Desktop (ex-Windsurf), Hermes, opencode, OpenClaw/ClawHub, and pi-mono; indexed on SkillsMP. The Installation section below has copy-paste setup for those plus Gemini CLI, GitHub Copilot, Cline, Trae, 通义灵码 Lingma, CodeBuddy/WorkBuddy, Qoder, Claude Desktop, and LM Studio. Desktop hosts that don’t auto-load skills (Claude Desktop, LM Studio) still get the tools over MCP — paste SKILL.md into the system prompt to add the routing layer.

SKILL.md’s frontmatter keeps only name, description, and license at the top level (all other fields are nested under metadata), so it validates cleanly under stricter skill-frontmatter parsers such as Codex’s.

Prerequisites

• An agent host with MCP support (Claude Code, Cursor, Codex, Gemini CLI, GitHub Copilot, Cline, Devin Desktop, Trae, 通义灵码 Lingma, CodeBuddy/WorkBuddy, Qoder, opencode, OpenClaw/ClawHub, pi-mono, etc.)

• An Asta API key — request here

  export ASTA_API_KEY=xxxxxxxxxxxxxxxx
  

Installation

Registering the MCP server is all you strictly need — Asta runs on any MCP-capable agent, and the tools work the moment the server connects. Step 2 adds an optional instruction layer (intent→tool routing, safe defaults, pitfall guards); the tools work fine without it.

Step 1: Register the Asta MCP Server

Every host points at the same endpoint — https://asta-tools.allen.ai/mcp/v1, streamable HTTP, authenticated with an x-api-key header. The only catch is that hosts name the fields differently (url vs serverUrl vs httpUrl, mcpServers vs servers), and a few can’t attach a custom header at all — those bridge through mcp-remote. Expand your agent:

Terminal / CLI

Editors & IDEs

IDE (China)

Desktop & local — MCP tools work once connected, but these hosts don’t auto-load skills, so paste SKILL.md in manually for the routing layer.

Step 2: Install the Skill (optional)

Skip this and the tools still work — this layer only sharpens tool selection and defaults. The skill body lives at skills/asta-skill/SKILL.md. The easiest path is the plugin marketplace.

Plugin marketplace (recommended)

# Any agent (Claude Code, Cursor, Copilot, etc.)
npx skills add Agents365-ai/365-skills -g

# Claude Code only
/plugin marketplace add Agents365-ai/365-skills
/plugin install asta

Also indexed on SkillsMP and ClawHub — each handles updates through its own marketplace.

Manual clone (any host)

git clone https://github.com/Agents365-ai/asta-skill.git /tmp/asta-skill
cp -r /tmp/asta-skill/skills/asta-skill /asta-skill

Verification

After registering the MCP server, installing the skill, and restarting your host, ask:

“Use Asta to get the paper ARXIV:1706.03762 with fields title,year,authors,venue,tldr”

A successful call returns Attention Is All You Need, NeurIPS 2017, Vaswani et al., with TLDR.

Usage

Just describe what you want:

> Use Asta to get the paper with DOI 10.48550/arXiv.1706.03762

> Search Asta for recent papers on mixture-of-experts at NeurIPS since 2023

> Who cited "Attention Is All You Need"? Show me the top 20 by citation count

> Find snippets in the Asta corpus that mention "flash attention latency"

> Look up Yann LeCun on Asta and list his 2024 papers

The skill picks the right Asta tool, attaches safe fields, and follows the documented workflow patterns.

Example: Search + Batch Download (chained with paper-fetch)

asta-skill only handles search and metadata; it does not download PDFs. To go from a search query to local PDFs, chain it with a paper-fetch skill (or any DOI-based downloader of your choice):

> Use Asta to find the 5 most cited papers on "single-cell ATAC-seq batch correction"
  since 2022, then hand the DOIs to paper-fetch to download all PDFs into ./papers/

What happens under the hood:

1. asta-skill → search_papers_by_relevance with publication_date_range="2022:" and fields=title,year,authors,venue,tldr,externalIds (note externalIds to expose DOI)
2. Agent extracts externalIds.DOI for each hit; falls back to externalIds.ArXiv when DOI is absent
3. paper-fetch → batch-resolves each DOI/arXiv ID through the Unpaywall → Semantic Scholar → arXiv → PubMed Central → bioRxiv/medRxiv → publisher-direct → Sci-Hub fallback chain
4. PDFs land in ./papers/, one per paper

paper-fetch is a separate skill — install it if you need download capability. asta-skill itself stays scoped to the Semantic Scholar corpus.

Step 1 — Asta returns the top 5 papers with DOIs:

Step 2 — paper-fetch downloads all 5 PDFs into ./papers/:

🔗 Related Skills

Part of the Agents365-ai research-skill family — pick the right tool for the job:

❤️ Support

If this skill helps you, consider supporting the author:

👤 Author

Agents365-ai

• GitHub:
• Bilibili:

📄 License

MIT