Papers By Ouroboros Apps

Papers by Ouroboros

Papers by Ouroboros 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.

• Works with ChatGPT, Claude, Gemini, Grok, Cursor, and any client that speaks Streamable HTTP plus OAuth 2.1 (dynamic client registration and PKCE)

Hosted server

• MCP server URL: https://papers-mcp.vercel.app/mcp (live, Streamable HTTP, OAuth sign-in)
• Health: https://papers-mcp.vercel.app/health
• Protected-resource metadata: https://papers-mcp.vercel.app/.well-known/oauth-protected-resource
• Connect guide: https://papers-mcp.vercel.app/connect
• Docs: https://ouroborosapps.com/docs/papers
• Registry name: io.github.LAHutchins91/papers (server.json points remotes[0].url at the hosted URL above)

If you self-host, the MCP endpoint is https://<your-host>/mcp. Set APP_BASE_URL to that origin and change remotes[0].url in your copy of server.json.

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://papers-mcp.vercel.app/mcp"
    }
  }
}

Claude Code:

claude mcp add --transport http papers https://papers-mcp.vercel.app/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.

Permissions, data, and limits

OAuth scope. One scope, papers. The consent screen names the assistant and that scope. It grants tool calls on this server only. It gives no access to any other account, file, or service.

Operations. All four tools are read-only lookups against public scholarly APIs (readOnlyHint: true, destructiveHint: false). There are no write or delete operations. Papers does not create, change, or remove anything on your behalf, outside the server, or in the scholarly sources.

Confirmations. Because nothing is written or deleted, no tool asks for a confirmation step. The only confirmations are the OAuth consent screen when you connect and Stripe Checkout when you start a trial.

Storage, retention, and privacy. Papers does not keep a log or history of queries. A search or identifier goes to the scholarly source that answers it and is not retained after the response. Papers stores a random account id (cookie 90 days, access token 1 hour, refresh token 30 days), single-use authorization code ids until they expire (5 minutes), and, after you start a trial, a subscription snapshot (account id, Stripe customer id, subscription id, status, period end, plan interval). Rate-limit counters are held in memory for about one to ten minutes. Connecting does not ask for your name or email. Export or deletion requests go to [email protected]. Full policy: https://papers-mcp.vercel.app/privacy

Free limits. Tool discovery (initialize, tools/list, ping) is open. Tool calls need a 14-day trial started through Stripe Checkout, then a Pro subscription. Stripe shows the amount; this repository does not list one. Each client IP is limited to 300 MCP requests per minute. The upstream APIs also apply their own rate limits (see Rate limits below).

How it is used. A typical session: the user asks a research question, the assistant calls search_papers, opens one or two results with get_paper, follows find_related_papers to see what cites them, and finishes with format_citation in the style the user needs. Every record in the answer links back to the source that returned it.

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 install
npm run dev

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

npm test
npm 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.

Contact

Public support, privacy, export, and deletion: [email protected]

• Privacy: /privacy
• Terms: /terms
• Support: /support
• Site: https://ouroborosapps.com

The public pages always show that mailbox. SCHOLARLY_CONTACT_EMAIL is sent only in scholarly API polite-use headers. It is not printed on the site. SUPPORT_EMAIL is not read.

server.json, glama.json, and .cursor-plugin/plugin.json have no contact-email field. The registry schema used by server.json does not define one.

Store-reviewer sign-in

Reviewers sign in on the OAuth consent screen with Reviewer sign-in. There is no separate sign-up, no second factor, and no card. Set these on the host (Vercel project → Settings → Environment Variables), for Production and for any Preview URL a reviewer will open. Do not commit the values.

Login checks the email and password hash. The allowlist is what skips the trial and Stripe. Set COMP_ACCOUNT_EMAILS to the same address as REVIEWER_LOGIN_EMAIL, or put the derived id in COMP_ACCOUNT_IDS. A normal anonymous account is unchanged and still needs a trial when billing is configured.

Generate the hash (the password is an argument, not a file in the repo):

node --input-type=module -e '
import crypto from "node:crypto";
const password = process.argv[1];
if (!password) { console.error("Pass the password as an argument."); process.exit(1); }
const salt = crypto.randomBytes(16);
const hash = crypto.scryptSync(password, salt, 32, { N: 16384, r: 8, p: 1 });
process.stdout.write(`scrypt$16384$8$1$${salt.toString("base64url")}$${hash.toString("base64url")}\n`);
' 'replace-with-the-reviewer-password'

Print the account id for COMP_ACCOUNT_IDS (first 32 hex characters of SHA-256 over papers-reviewer, a NUL byte, and the lowercased email):

node --input-type=module -e '
import crypto from "node:crypto";
const email = process.argv[1].trim().toLowerCase();
process.stdout.write(crypto.createHash("sha256").update(`papers-reviewer\0${email}`).digest("hex").slice(0, 32) + "\n");
' '[email protected]'

After sign-in, the tools return live records. A reviewer can call search_papers with a short query, then get_paper, find_related_papers, and format_citation with DOI 10.1038/nature14539.

Papers by Ouroboros. More from Ouroboros: https://ouroborosapps.com