Asta MCP — Academic Paper Search

Asta is Ai2's Scientific Corpus Tool, exposing the Semantic Scholar academic graph over MCP (streamable HTTP transport). This skill tells agents which Asta tool to call for which intent, and how to compose them into useful workflows.

• MCP endpoint: https://asta-tools.allen.ai/mcp/v1
• Auth: x-api-key header (request key at https://share.hsforms.com/1L4hUh20oT3mu8iXJQMV77w3ioxm)
• Transport: streamable HTTP

Prerequisite Check

Before invoking any tool, verify the Asta MCP server is registered in the host agent. Tool names will be prefixed by the MCP server name chosen at install time (commonly asta__<tool> or mcp__asta__<tool>). If no Asta tools are visible, direct the user to the Installation section below.

Tool Map — Intent → Asta Tool

Search/citation tools accept publication_date_range (format YYYY-MM-DD:YYYY-MM-DD; year shorthand like "2021:", ":2015-01", "2015:2020" is also accepted) and venues (comma-separated) filters, plus fields for field selection — pass them whenever the user's intent constrains scope (e.g., "recent", "since 2022", "at NeurIPS").

⚠️ fields parameter — avoid context blowups

get_paper / get_paper_batch accept a fields string. Never request citations or references via fields — a single highly-cited paper (e.g. Attention Is All You Need) returns 200k+ characters and will overflow the agent's context window. Use the dedicated get_citations tool for forward citations (it paginates). Asta does not provide a dedicated get_references tool — to retrieve a paper's reference list, use get_paper with fields=references only for papers you know have a small reference list (typically < 100).

Safe default fields for get_paper:

title,year,authors,venue,tldr,url,abstract

Add journal, publicationDate, fieldsOfStudy, isOpenAccess only when needed.

Retrieving DOI / external IDs (undocumented but supported)

Asta's official fields list does not include externalIds, but the field is transparently passed through to the underlying Semantic Scholar API and works in practice. Add externalIds to fields to retrieve DOI, PubMed, PubMedCentral, ArXiv, MAG, DBLP, CorpusId. Caveats:

• Not all papers have a DOI — pure arXiv preprints often only return ArXiv + CorpusId.
• get_paper("DOI:...") lookup is not 100% reliable; some valid DOIs return not found. Prefer searching by title first, then reading externalIds off the result.
• Since this is undocumented, treat it as best-effort and degrade gracefully if a future Asta release drops it.

Workflow Patterns

Pattern 1 — Topic Discovery

1. search_papers_by_relevance(keyword, publication_date_range="<current_year-5>:", venues=?) → initial hits (compute the lower bound from today's date — e.g., in 2026 pass publication_date_range="2021:"; adjust or drop the filter if the user asks for older work)
2. Rank/present top N by citationCount + recency
3. Offer follow-ups: get_citations on the most influential, or snippet_search for specific claims

Pattern 2 — Seed-Paper Expansion

1. get_paper(DOI|arXiv|...) → verify seed
2. get_citations(paperId) → forward expansion
3. Optionally search_papers_by_relevance with seed title terms for sideways discovery
4. Deduplicate by paperId before presenting

Pattern 3 — Author Deep-Dive

1. search_authors_by_name(name) → pick correct profile (disambiguate by affiliation)
2. get_author_papers(authorId) → full publication list
3. Filter client-side by topic keywords or date

Pattern 4 — Evidence Retrieval

1. snippet_search(claim_query) → find passages making/supporting a claim
2. For each hit, optionally get_paper(id) for full metadata

Output & Interaction Rules

• Always report total count and which tool was used.
• Present top 10 as a table (title, year, venue, citations), then details for the most relevant.
• If the user writes in Chinese, present summaries in Chinese; keep titles in original language.
• After results, offer: Details / Refine / Citations / Snippet / Export / Done.

Critical Rules

• Prefer batched intent over ping-pong. If the user's question needs two independent lookups, issue them as parallel MCP tool calls in one turn, not sequentially.
• Never guess IDs. If a user gives a fuzzy title, use search_paper_by_title before get_paper.
• Respect rate limits. An API key buys higher limits but not unlimited — stop expanding citation graphs beyond what the user asked for.
• Do not fabricate fields. If Asta returns null abstract or venue, say so rather than inventing.

Handling Asta responses