Papers

Papers by Ouroboros

Papers is a remote MCP server for students, researchers, writers, and clinicians who want an assistant to search and cite real papers. It is a lighter, independent alternative to Consensus, SciSpace, and Elicit.

The assistant can search OpenAlex, Semantic Scholar, PubMed, Crossref, and arXiv, open a paper by DOI, PMID, or arXiv id, see what cites a paper or what that paper cites, and format APA, MLA, Chicago, or BibTeX. Every result includes a source link and an identifier that came back from one of those APIs. If an API returns nothing, Papers returns nothing. It does not fill gaps with invented citations.

• MCP endpoint: https://<your-host>/mcp
• Works with ChatGPT, Claude, Gemini, Grok, Cursor, and any client that speaks Streamable HTTP plus OAuth 2.1 (dynamic client registration and PKCE)

server.json is the MCP Registry manifest (io.github.LAHutchins91/papers). Its remotes[0].url is a placeholder (https://papers-mcp.vercel.app/mcp). Change it to the origin you actually deploy, and set APP_BASE_URL to that same origin.

Connect

Leave the client id and secret empty so the client can register itself.

Cursor, in ~/.cursor/mcp.json or a project .cursor/mcp.json:

{  "mcpServers": {    "papers": {      "url": "https://<your-host>/mcp"    }  }}

Claude Code:

claude mcp add --transport http papers https://<your-host>/mcp

Other clients: add the same URL and choose OAuth. The consent screen names the assistant and the papers scope. Tool discovery (initialize, tools/list, ping) does not require a token. Calling a tool does.

Tools

• search_papers — question or keywords, with optional year range, field, open-access flag, and source (openalex, semantic_scholar, pubmed, crossref, arxiv, or all)
• get_paper — one record and its abstract, by DOI, PMID, arXiv id, or OpenAlex work id
• find_related_papers — cited_by or references for that identifier
• format_citation — apa, mla, chicago, or bibtex from the fetched metadata

Titles are kept as the source wrote them. Author names are split from display names, so particles such as “van” can land in the wrong part and deserve a look before you publish the citation.

Trial and billing

Start a 14-day trial from the account page after you connect an assistant. Stripe Checkout is the only place an amount is shown. Monthly and yearly plans use the price ids you configure. Papers does not create Stripe products.

When STRIPE_SECRET_KEY, STRIPE_PRICE_MONTHLY, and STRIPE_PRICE_YEARLY are all set, tool calls require a Stripe subscription in trialing or active status. When they are not set, GET /health reports billingConfigured: false and an OAuth-signed caller can use the tools. Set the Stripe variables before you expose a deployment publicly.

Storage

Paper search is stateless. Subscription snapshots and used authorization-code ids share one store:

The document is { accounts, usedCodes }. Each account is { userId, stripeCustomerId, subscriptionId, subscriptionStatus, currentPeriodEnd, plan, updatedAt }. usedCodes maps an authorization-code id to its expiry, in unix seconds. Expired ids are dropped on every write. An older file that is only a map of accounts is still read. There is no separate database schema. When billing is configured, Stripe is the source of truth: a cold process looks the customer up by metadata.papers_user_id if the snapshot is missing.

OAuth access tokens and client registrations are signed tokens, not rows. Set AUTH_SIGNING_SECRET in production so tokens survive a restart. On Vercel, set STORAGE_BACKEND=blob before the deployment is public.

Environment

Do not commit secrets. Copy what you need into the host’s environment, not into the repo.

Run locally

npm installnpm run dev

Open http://127.0.0.1:43127. Health is GET /health.

npm testnpm run typecheck

Tests call the public APIs that do not need keys. Stripe is not called. Semantic Scholar often answers 429 from shared IP space; the tool reports that rate limit and does not substitute a made-up paper.

Deploy

Vercel: vercel.json builds api/index.ts and routes every path to that function, including logo.jpg in the function bundle. Repository files such as /package.json and /src/* are not served as static files. Set the environment variables above, set APP_BASE_URL to the deployment origin, set STORAGE_BACKEND=blob, and update server.json remotes[0].url to https://<that-host>/mcp. Create the Blob store in the Vercel project and leave its read-write token in BLOB_READ_WRITE_TOKEN. Stripe’s webhook endpoint is POST /billing/webhook.

Docker:

docker build -t papers-mcp .docker run --rm -p 43127:43127 -e APP_BASE_URL=http://127.0.0.1:43127 -e AUTH_SIGNING_SECRET=replace-me -e [email protected] papers-mcp

The container listens on 43127.

Rate limits

Requests are spaced per host: OpenAlex about 10/s, Crossref about 4/s, PubMed about 3/s without an NCBI key, Semantic Scholar about 1/s without a key, and arXiv at least 3 seconds between calls. A descriptive User-Agent is always sent. The contact email is included only when SCHOLARLY_CONTACT_EMAIL is set.

License

MIT. Copyright Lawrence Hutchins. See LICENSE.