lune-mcp-server
Official MCP server for Lune Research: search top-tier papers and methodology guidance from any AI client.
Documentation
@retrograde-labs/lune-mcp-server
Official Model Context Protocol server for Lune Research.
Exposes 12 tools and 6 guided research workflows (prompts) for searching, retrieving, comparing, and fact-checking academic papers across security, ML, NLP, CV, and systems venues, plus retrieval over your own uploaded workspace documents. Two transports:
- stdio: run locally via `npx @retrograde-labs/lune-mcp-server`. Reads `LUNE_API_KEY` from the environment.
- Streamable HTTP: hosted at `https://mcp.luneresearch.com`. Pass your PAT or OAuth token as `Authorization: Bearer ...`.
Quick start (Claude Desktop, Cursor, etc.)
{
"mcpServers": {
"lune-research": {
"command": "npx",
"args": ["-y", "@retrograde-labs/lune-mcp-server"],
"env": {
"LUNE_API_KEY": "lune_your_personal_access_token"
}
}
}
}Get your token at https://luneresearch.com/dashboard/settings/credentials.
Tools
| Tool | Description |
|---|---|
| `search_papers` | Hybrid vector + BM25 search across the corpus |
| `search_papers_many` | Run many query variants in one call, RRF-merged |
| `search_related_papers` | Semantically nearest papers to a given paper |
| `get_paper_fulltext` | Parsed full text (markdown or JSON) |
| `get_paper_citations` | Citation graph (cited_by or cites) |
| `list_conferences` | Indexed venues, optionally by category |
| `get_conference_papers` | Paginated papers for a venue |
| `extract_from_papers` | Structured field extraction across many papers |
| `verify_claims` | Fact-check claims against the corpus with quotes |
| `gather_evidence` | Judge evidence sufficiency for a task, with gaps + next queries |
| `search_research_guidance` | Curated reproducibility / methodology corpus |
| `get_research_guidance_doc` | Full text of a guidance document |
Prompts
Reusable research workflows, surfaced by MCP clients as slash commands (e.g. `/literature_review`). Each runs a guided, multi-tool sequence grounded in the corpus, so common research tasks are one command instead of hand-orchestrating the tools.
| Prompt | What it does | Key arguments |
|---|---|---|
| `/literature_review` | Survey a topic and synthesise themes, foundational vs recent work, and open gaps | `topic` (+ optional `venues`, `since_year`) |
| `/find_related_work` | From your abstract, find and organise prior work to cite and distinguish your contribution from | `abstract` (+ optional `venues`) |
| `/compare_papers` | Build a structured comparison table across papers, read from full text | `topic` (+ optional `columns`) |
| `/verify_draft` | Fact-check a draft or list of claims against the corpus, with a verbatim quote per claim | `draft` |
| `/trace_citations` | Trace a paper's lineage: its foundations, what built on it, and adjacent work | `paper` |
| `/research_methodology` | Grounded advice on experiment design, ablations, evaluation, rebuttals, or venue choice | `question` |
Debugging with MCP Inspector
The repository pins MCP Inspector 2.3.0 and checks in a read-only server
configuration for both protocol eras. Node 24 from `.nvmrc` satisfies the
Inspector requirement.
With a PAT in `LUNE_API_KEY`, this builds the stdio server and opens the
Inspector catalog with the PAT available to local processes. Select
`lune-stdio-modern` to exercise 2026-07-28:
export LUNE_API_KEY=lune_xxxxxxxxxxxxxxxxxxxxxxxx
bun run mcp:inspectWith no `LUNE_API_KEY`, the same command opens the server catalog without an
environment override. Select `lune-remote-modern`; Inspector starts OAuth when
you connect. The launcher never passes an empty `-e` assignment.
The printed URL contains a one-time Inspector token. Open that URL instead of
typing the port manually. The Inspector stays on loopback with authentication
enabled.
The hosted-server mnemonic opens the same catalog without building stdio:
bun run mcp:inspect:remoteThe web, CLI, and TUI clients share OAuth state on disk. After signing in once,
the CLI can check the hosted server without opening another browser:
bun run mcp:inspect:remote:checkThe local stdio checks use a non-production test token and a deliberately dead
API URL. They validate connection, version negotiation, and the tool catalog
without consuming quota:
bun run mcp:inspect:check
bun run mcp:inspect:check:legacy
bun run mcp:inspect:tuiThe local and remote configs at `apps/mcp/inspector.*.config.json` pin
`protocolEra` explicitly. Inspector defaults to the legacy era, so an
unconfigured launch does not prove 2026-07-28 behavior. During debugging, pin
Protocol beside Console for stdio or Network for HTTP. Test invalid tool
arguments, missing prompt arguments, both protocol eras, and expired OAuth
tokens as well as successful calls.
Run the frozen official 2026-07-28 server requirements with:
bun run mcp:conformanceThe conformance harness binds only `127.0.0.1` and injects a local test token.
Its baseline names individual checks for diagnostic capabilities Lune does not
advertise, such as image tools and readable resources. Every applicable check
still gates the run, and a stale baseline entry fails when support is added.
Never write diagnostics to stdout on stdio. The server and its cache write
diagnostics to stderr so JSON-RPC framing stays intact.
License
MIT
Frequently asked questions
What is lune-mcp-server?
lune-mcp-server is Official MCP server for Lune Research: search top-tier papers and methodology guidance from any AI client.
How do I install lune-mcp-server?
Open the GitHub repository and follow its README. Most MCP servers are added to your client's MCP config, then called by your agent.
Is lune-mcp-server open source?
Yes — it is hosted on GitHub at https://github.com/RetrogradeLabs/lune-mcp-server and has 4 stars.
Related MCP tools
🔌 Plug-and-play library to enable agents to call MCP and UTCP tools via code execution.
Code research platform for AI agents; find, understand, and prove context across your code and all of GitHub, in a fraction of the tokens. One toolset, MCP or CLI
A Model Context Protocol (MCP) server and CLI that provides tools for agent use when working on iOS and macOS projects.
The go-to web for your AI coding agent — local-first search, fetch, crawl & research over MCP. No API keys, no cloud, $0/query. Public beta.
Mobile app automation and verification for AI coding agents. CLI, MCP server, and typed Node.js API for iOS, Android, HarmonyOS, TV, web, macOS, and Linux.
MCP Aggregator, Orchestrator, Middleware, Gateway in one docker
Run your own MCP server? See who uses it and what to fix.
Measure it with TrackMCP