@retrograde-labs/lune-mcp-server

The official Lune Research MCP server. It lets an AI agent search the full text of papers from top-tier venues, follow citations, and look up curated research-methodology guidance, so its answers can cite a paper's title, authors and venue instead of relying on memory.

Install

Hosted (recommended). Add this URL to any client that speaks Streamable HTTP:

https://mcp.luneresearch.com

A client that supports MCP authorization signs you in through the browser, so there is no token to paste or rotate. In Claude Code:

claude mcp add --transport http --scope user lune https://mcp.luneresearch.com

Then run /mcp inside Claude Code, choose lune, and sign in through the browser tab that opens.

A client without OAuth support sends a Lune access key in the Authorization header instead. Create one on the Credentials page: open API keys and click New key. This request lists the tools, which makes it a quick way to check a key:

curl https://mcp.luneresearch.com \
  -H "Authorization: Bearer lune_your_access_key" \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}'

Setup notes for specific apps are in the docs: web AI apps such as ChatGPT, Claude and Perplexity, and local AI agents such as Cursor, VS Code and Codex.

Local stdio. For clients that only speak stdio, run the package directly and hand it an access key. It needs Node.js 22 or newer:

{
  "mcpServers": {
    "lune": {
      "command": "npx",
      "args": ["-y", "@retrograde-labs/lune-mcp-server"],
      "env": { "LUNE_API_KEY": "lune_your_access_key" }
    }
  }
}

The server reads LUNE_API_KEY from the environment and writes diagnostics to stderr, never stdout, so JSON-RPC framing stays intact.

Tools

Twelve tools. Most follow a prefix: search_* finds candidates, get_* retrieves something you already have a handle on, and list_* browses. The hosted endpoint also offers get_more_tools, which an agent calls to tell Lune about a capability it needed and could not find. The stdio package does not include it.

Finding papers

Tool	What it does
search_papers	Hybrid vector + BM25 search over the corpus
search_papers_many	Many query variants in one call, merged by reciprocal rank fusion
search_related_papers	The nearest neighbors of a paper you already have
list_conferences	Indexed venues, optionally filtered by category
get_conference_papers	One venue's papers, paginated

Reading them

Tool	What it does
get_paper_fulltext	Parsed full text, as markdown or JSON
get_paper_citations	The citation graph, in either direction

Working across many at once

Tool	What it does
extract_from_papers	Pull the same structured fields out of a set of papers
verify_claims	Check each claim against the corpus, with a verbatim quote when one exists
gather_evidence	Judge whether the evidence is sufficient yet, and name what is missing

Research methodology

Tool	What it does
search_research_guidance	The curated methodology and reproducibility corpus
get_research_guidance_doc	The full text of one guidance document

Prompts

Six prompts, which many MCP clients list as slash commands. Each one hands the agent a brief for a common research task and names the tools to call, so the task takes one command instead of a plan you have to spell out.

Prompt	What it does	Arguments
/literature_review	Survey a topic: the main themes, foundational and recent work, and open gaps	topic (+ venues, since_year)
/find_related_work	Read your abstract, then find the prior work to cite and distinguish yourself from	abstract (+ venues)
/compare_papers	Build a comparison table across papers, read out of their full text	topic (+ columns)
/verify_draft	Check each claim in a draft, with a verbatim quote when one exists	draft
/trace_citations	Trace a paper's lineage: what it stands on, what stands on it	paper
/research_methodology	Grounded advice on experiment design, ablations, evaluation, rebuttals, venues	question

Authentication

The hosted endpoint takes one credential on every request, as Authorization: Bearer <credential>. The stdio package reads an access key from LUNE_API_KEY instead. The credential is one of these:

• An OAuth access token, which an MCP client obtains by signing you in through the browser. Hosted endpoint only.
• A Lune access key (lune_...), created on the Credentials page under API keys. Works on both transports.

An OAuth connection always bills the signed-in person's personal team, drawing on its plan, daily allowance and credits. An access key bills the team it was created in, so a key is how a shared team's plan pays for an agent. How requests are counted is in quotas and billing.

OAuth for connector builders

Most MCP clients run this flow on their own. If you are building a connector, the details follow. auth.md walks through each request with curl.

An unauthenticated POST https://mcp.luneresearch.com answers 401 with this challenge:

WWW-Authenticate: Bearer resource_metadata="https://mcp.luneresearch.com/.well-known/oauth-protected-resource", scope="papers:read guidance:read account:read"

The protected-resource document it points to names the authorization server in authorization_servers[0]. That server is https://api.luneresearch.com, the authorization server for MCP connectors. It publishes RFC 8414 metadata at https://api.luneresearch.com/.well-known/oauth-authorization-server and has no OpenID Connect discovery document.

Endpoint	URL
Issuer	https://api.luneresearch.com
Authorization	https://api.luneresearch.com/oauth/authorize
Token	https://api.luneresearch.com/oauth/token
Client registration (RFC 7591)	https://api.luneresearch.com/oauth/register
Revocation (RFC 7009)	https://api.luneresearch.com/oauth/revoke
JWKS	https://api.luneresearch.com/.well-known/jwks.json
Authorization server metadata	https://api.luneresearch.com/.well-known/oauth-authorization-server
Protected resource metadata	https://mcp.luneresearch.com/.well-known/oauth-protected-resource

The flow:

1. Register a client. Dynamic client registration is the only way to get a client_id; there are no pre-registered client IDs. Clients are public, so there is no secret and the token endpoint auth method is none. A platform that cannot register on the fly can do it once by hand and keep the client_id:

   curl -X POST https://api.luneresearch.com/oauth/register \
     -H "Content-Type: application/json" \
     -d '{"client_name": "My connector", "redirect_uris": ["https://example.com/oauth/callback"]}'
   

2. Send the person to the authorization endpoint with response_type=code, client_id, redirect_uri, scope, state, code_challenge, code_challenge_method=S256 and resource. PKCE with S256 is required, and so are state and scope.

3. Handle the callback. After the person signs in and approves, Lune redirects to your redirect_uri with code, state and iss=https://api.luneresearch.com (RFC 9207). Check that state is the one you sent and that iss equals the issuer. If the person declines, the redirect carries error=access_denied with the same state and iss.

4. Exchange the code. POST grant_type=authorization_code, code, redirect_uri, client_id, code_verifier and resource to the token endpoint as application/x-www-form-urlencoded.

5. Call the server with Authorization: Bearer <access_token>. Once the token expires the server answers 401 with error="invalid_token". To refresh, POST grant_type=refresh_token and refresh_token to the token endpoint.

Requirements and limits:

• Request exactly papers:read guidance:read account:read, the scopes the protected-resource document lists. Together they cover every tool.
• Send resource=https://mcp.luneresearch.com on both the authorization request and the token request. Without it the access token is bound to your client_id instead of the server, and requests on protocol version 2026-07-28 fail with 401 invalid_token, which the person sees as a sign-in loop. Older protocol versions still accept such a token, but only for a transition period.
• Redirect URIs must use https, http on 127.0.0.1, [::1] or localhost, or a private-use scheme such as cursor://, and must not contain a fragment. The authorization request must repeat a registered URI exactly, port included.
• An authorization code expires after 10 minutes and works once; a failed exchange uses it up. Access tokens are RS256 JWTs that expire after one hour (expires_in: 3600).
• Every refresh returns a new refresh token. Store it and drop the old one. Each refresh token is valid for 90 days, so a connection in regular use does not expire. Presenting a refresh token more than 30 seconds after it was replaced revokes every refresh token in that chain, and the person has to connect again. Access tokens already issued keep working until they expire.
• To disconnect, POST token=<refresh_token> (form-encoded) to the revocation endpoint.

Scopes

papers:read covers the paper corpus, full text and citations, guidance:read covers the research-guidance corpus, and account:read reads your identity and usage.

A credential that lacks the scope a tool needs fails differently depending on its type. An OAuth token gets HTTP 403 with a challenge that names only the missing scope, so the client can ask the person to approve it:

WWW-Authenticate: Bearer error="insufficient_scope", error_description="The papers:read scope is required for this request.", resource_metadata="https://mcp.luneresearch.com/.well-known/oauth-protected-resource", scope="papers:read"

An access key gets a tool error that names the missing scope and tells the agent to have you create a key that includes it.

Revoking access

Revoke access keys under API keys and connected apps under OAuth clients on the Credentials page. OAuth connections are listed under your personal team, so select it first. A revoked access key stops working within five minutes. Revoking an app cancels its refresh tokens at once, but the access token it already holds keeps working until it expires, at most an hour later.

Transports

Transport	Where	MCP protocol versions
Streamable HTTP	https://mcp.luneresearch.com	2026-07-28, 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05, 2024-10-07
stdio	npx -y @retrograde-labs/lune-mcp-server	2026-07-28, 2025-11-25, 2025-06-18, 2025-03-26, 2024-11-05, 2024-10-07

About the hosted endpoint:

• /mcp and /v1/mcp answer too, for older installs. Discovery follows the path, so a client on /mcp is pointed at https://mcp.luneresearch.com/.well-known/oauth-protected-resource/mcp. A token bound to any of the three URLs works on all three.
• Requests are stateless. Each one carries its own credential and there is no session ID, so a client that has been idle for hours keeps working without re-initializing.
• MCP traffic is POST only. GET and DELETE return 405, so there is no SSE stream to open, and the server sends no notifications of its own. A browser that opens the bare URL is redirected to the docs instead.
• Send Accept: application/json, text/event-stream. On older protocol versions, a request that leaves out either type gets 406.
• A request with an Origin header gets 403 unless that origin is on the server's allowlist, so connect from a server or a native client rather than from a web page.
• JSON-RPC batches, which only older protocol versions use, may hold up to 50 messages. A request body may be up to 1 MB.
• The Registry manifest is served at https://mcp.luneresearch.com/.well-known/mcp/server.json.

Links

• Connection details
• Web AI apps
• Local AI agents
• Authentication guide for agents (auth.md)
• Quotas and billing
• Troubleshooting

Support

Bug reports and feature requests go to the issue tracker. For account questions, email support@luneresearch.com. To report a security issue, follow SECURITY.md.

License

MIT