acatome-quest-mcp

Paper-request MCP for scientific papers. The missing piece between precis-mcp (navigates what's already in your library) and acatome-extract (ingests PDFs that land in an inbox).

An LLM says "I want this paper" (DOI, arXiv id, title, or free-form citation). Quest:

1. Checks the store first — no duplicate work if we already have it.
2. Resolves the metadata via Crossref + Semantic Scholar + arXiv.
3. Flags misconceptions — broken DOI, DOI↔title mismatch, duplicate of an existing slug, fabrication suspect.
4. Fetches the PDF from legitimate open-access sources only and drops it into the existing watch inbox, where acatome-extract takes over.
5. Returns a request id in milliseconds. Slow extraction happens out of band; the MCP call never blocks.

Open access only — by policy

Quest fetches from arXiv, Unpaywall, OpenAlex, Europe PMC, and Semantic Scholar's open-access index only. It does not, will not, and cannot be configured to use Sci-Hub, LibGen, institutional proxies without explicit opt- in, or any other paywall-circumvention mechanism. Failed retrievals yield a needs_user status with the publisher URL, for you to retrieve manually.

Install

pip install acatome-quest-mcp
# or with uv
uv add acatome-quest-mcp

For dedup against a local acatome-store:

pip install 'acatome-quest-mcp[store]'

Four tools

submit

submit(ref={"doi": "10.1021/jacs.2c01234"})
submit(ref={"title": "Anion exchange membranes for NOx reduction",
            "authors": ["Feng, Z."], "year": 2024})
submit(ref={"raw": "Feng et al. 2024, Adv. Funct. Mater. 34, 2300512"})
submit(ref={"doi": "10.1234/x"}, dry_run=True)      # resolve only, no queue
submit(ref={"doi": "10.1234/x"},
       source={"document": "ch02.tex", "line": 147})

Response:

{
  "id": "9f3b…",
  "status": "found_in_store",
  "resolved": {"doi": "10.1021/jacs.2c01234",
               "title": "…", "authors": ["…"], "year": 2024,
               "ref": "smith2022jacs"},
  "candidates": [],
  "misconceptions": []
}

status

status(id="9f3b…")
status(filter={"status": "needs_user"})
status(filter={"created_by": "asa", "has_misconception": True})
status(filter={"source_document": "ch02.tex"})

update

update(id, mode="confirm", choice=0)            # pick candidates[0]
update(id, mode="repoint", doi="10.1023/A:…")   # user-corrected DOI
update(id, mode="flag", code="retracted",
       evidence="Retraction Watch 2024-08-12")
update(id, mode="priority", priority=5)
update(id, mode="cancel")

submit_file

# User drops a PDF for an already-tracked request (reopens failed / needs_user):
submit_file(url="https://cdn.discordapp.com/…/paper.pdf",
            request_id="7f3a…",
            filename="feng2024.pdf")

# User supplies both a PDF and a DOI in one step (creates the request):
submit_file(url="https://cdn.discordapp.com/…/paper.pdf",
            ref={"doi": "10.1021/jacs.2c01234"},
            created_by="asa")

# Bytes already in memory (no URL to fetch):
submit_file(content_base64="JVBERi0xLjQKJf…",
            request_id="7f3a…")

PDF magic bytes are validated; HTML error pages are rejected. The file is written to the extractor's inbox (~/.acatome/inbox/ by default) and the request flips to ingesting. If the paper's DOI is already in the store, the tool short-circuits to found_in_store without writing anything.

CLI

The acatome-quest binary exposes the same surface as the MCP plus a couple of shell-friendly helpers:

acatome-quest submit 10.1021/jacs.2c01234
acatome-quest status <id>
acatome-quest status --filter status=needs_user
acatome-quest status --filter status=needs_user --count     # just prints "3"
acatome-quest update <id> repoint --doi 10.1023/A:…
acatome-quest submit-file --path ./feng2024.pdf --request-id 7f3a…
acatome-quest submit-file --url https://.../paper.pdf --doi 10.1021/jacs.2c01234
acatome-quest report                                         # markdown worklist
acatome-quest report --document ch04.tex --format markdown   # scoped
acatome-quest runner [--once]
acatome-quest reconcile

report renders a paste-ready markdown document for every request in needs_user, failed, or extract_failed — each entry with citation, DOI/arXiv link, failure reason, misconception evidence, and a concrete suggested action (repoint DOI, drop PDF into ~/.acatome/inbox/, request via interlibrary loan, …). Hand it to a librarian or paste into an ILL form.

Statuses

Misconception codes

Architecture

 agent ──submit()──► acatome-quest-mcp (FastMCP, stdio)
                            │
                            ▼
              cluster.papers.requests (Postgres)
                            │
                            ▼
           acatome-quest-runner (launchd, poll 30 s)
                  │
          fetch: arxiv → unpaywall → …
                  │
                  ▼
          ~/.acatome/inbox/<slug>__<hash>.pdf
                  │
                  ▼
          acatome-extract watch  →  acatome-store
                  ▲
                  └── runner polls by DOI, flips to `ingested`

Configuration

Development

uv sync
uv run pytest
uv run ruff check .
uv run mypy src tests

License

GPL-3.0-or-later. See LICENSE.