confer MCP Server

A Model Context Protocol stdio server that exposes the confer paper corpus to AI agents — Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Codex, and any other MCP-capable client.

Tools

| Tool | Description | |------|-------------| | list_venues | List all conferences/journals, optionally filter by category or series. | | search_papers | Field-aware full-text search (author:, title:, inst:, track:, venue:, year:, keyword:, doi:; - prefix excludes). | | get_paper | Retrieve the full record for a paper by venue + id. | | find_similar | TF-IDF cosine-similarity neighbours across the full corpus. | | top_authors | Most prolific authors (optionally filtered by venue or query). | | top_institutions | Most represented institutions. | | top_tracks | Most common tracks/topics. | | export_bibtex | BibTeX export for a list of {venue, id} references. |

Quick start (no clone, no build)

The server is published on npm as confer-mcp. Most clients use the same block below — npx fetches and runs the server, which streams the corpus from https://confer.repus.me/data on demand (no local data needed):

{
  "mcpServers": {
    "confer": {
      "command": "npx",
      "args": ["-y", "confer-mcp"]
    }
  }
}

Requires Node ≥ 18 on your PATH. Restart the client after editing its config.

The first query fetches data and may take a few seconds; later queries are fast.

Set it up in your client

Claude Desktop

Edit the config file, add the mcpServers block above, and restart Claude Desktop.

• macOS — ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows — %APPDATA%\Claude\claude_desktop_config.json

Claude Code (CLI)

One command — no file editing:

claude mcp add confer -- npx -y confer-mcp
# user-wide instead of project-local:
claude mcp add confer -s user -- npx -y confer-mcp

Cursor

Add the mcpServers block to ~/.cursor/mcp.json (global) or .cursor/mcp.json (per-project), or use Settings → MCP → Add new server.

VS Code (GitHub Copilot)

VS Code uses the key servers (not mcpServers). Add to .vscode/mcp.json (workspace) or your user mcp.json, then enable it from the Copilot chat Tools picker — or run MCP: Add Server from the Command Palette.

{
  "servers": {
    "confer": {
      "command": "npx",
      "args": ["-y", "confer-mcp"]
    }
  }
}

Cline / Windsurf / other VS Code agents

Use the standard mcpServers block above — Cline in its cline_mcp_settings.json, Windsurf in ~/.codeium/windsurf/mcp_config.json (or each tool's "Add MCP server" UI).

Codex CLI (OpenAI)

Codex uses TOML at ~/.codex/config.toml:

[mcp_servers.confer]
command = "npx"
args = ["-y", "confer-mcp"]

Or add it from the CLI:

codex mcp add confer -- npx -y confer-mcp

ChatGPT (remote connectors only)

ChatGPT's Developer Mode connectors accept a public HTTPS MCP URL — they can't launch a local npx command. To use confer in ChatGPT you must expose this stdio server over HTTP and make it publicly reachable, e.g. bridge it with a stdio→HTTP gateway and a tunnel:

# 1. bridge the stdio server to a local HTTP/SSE endpoint
npx -y supergateway --stdio "npx -y confer-mcp"
# 2. expose that port publicly (separate terminal)
ngrok http <port-printed-by-supergateway>

Then in ChatGPT: Settings → Connectors → Advanced → Developer mode → Add custom connector, and paste the public https://….ngrok.app/mcp URL. See the supergateway and ChatGPT connector docs for exact flags. (This keeps a process + tunnel running on your machine.)

Configuration

| Env var | Default | Description | |---------|---------|-------------| | CONFER_DATA_URL | https://confer.repus.me/data | Base URL the server fetches venues.json and per-venue JSON from. | | CONFER_DATA_DIR | (auto) | Read from a local directory instead of the network. Auto-detected when run from a repo checkout (web/public/data); set it to force a path and run fully offline. | | HTTPS_PROXY / HTTP_PROXY | — | Used for the remote fetch when set (e.g. behind a corporate proxy). |

Run from a local checkout (offline / development)

cd mcp
npm install
npm run build      # → dist/server.js

Then point your client at the built file (it auto-uses the repo's web/public/data, so it works offline):

{
  "mcpServers": {
    "confer": {
      "command": "node",
      "args": ["/absolute/path/to/confcrawl/mcp/dist/server.js"]
    }
  }
}

Example queries

list_venues(series="ICSE")
search_papers(query="author:lecun deep learning", venues=["icml2025","iclr2026"], limit=10)
get_paper(venue="icse2025", id="<paper-id>")
find_similar(venue="icse2025", id="<paper-id>", n=10)
top_authors(venues=["fse2025","fse2026"], limit=15)
top_institutions(query="venue:ICSE year:2025")
export_bibtex(refs=[{"venue":"icse2025","id":"<id>"},{"venue":"fse2025","id":"<id>"}])

Query syntax

| Prefix | Matches | |--------|---------| | author:smith | Author name contains "smith" | | title:"code review" | Title contains the phrase | | inst:mit | Author institution contains "mit" | | track:testing | Track name contains "testing" | | venue:ICSE | Venue name or id contains "ICSE" | | year:2025 | Venue year is 2025 | | keyword:fuzzing | Paper keyword contains "fuzzing" | | -author:doe | Exclude papers where author contains "doe" | | (bare word) | Matches anywhere in title, abstract, authors, tracks, etc. |

Multiple terms are combined with AND semantics.