Model Context Protocol

Connect Claude to the directory

The AI Safety Research Directory exposes its curated, AI-classified paper dataset as a public MCP server. Any MCP-aware client — Claude Desktop, Claude Code, Cursor, your own agent — can connect and query the data inside a conversation. No API key. Refreshes nightly.

Server URL
https://aisafety.anthroforge.io/api/mcp
Transport
Streamable HTTP
Auth
None (read-only public)
Refresh
Nightly, 00:30 UTC

Claude Desktop

Edit ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows), then fully quit and relaunch Claude Desktop.

{
  "mcpServers": {
    "aisafety": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://aisafety.anthroforge.io/api/mcp"]
    }
  }
}

Uses the mcp-remote adapter to bridge Claude Desktop's stdio transport to our streamable HTTP endpoint. Works across all Claude Desktop versions that support MCP. Requires Node.js on your machine.

Claude Code

One command in your terminal — the server becomes available in any Claude Code session.

claude mcp add --transport http aisafety https://aisafety.anthroforge.io/api/mcp

MCP Inspector

For exploring the server interactively — useful when building your own MCP client.

npx @modelcontextprotocol/inspector https://aisafety.anthroforge.io/api/mcp

Quick smoke test

Verify the server is up without any MCP client installed.

curl -sf -X POST https://aisafety.anthroforge.io/api/mcp \
  -H "Content-Type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/list"}' | jq .

Available tools

search_papers

Search AI safety research papers by free-text query, tags, and minimum safety score. Returns up to `limit` papers ordered by safety score then publication date.

Arguments
querystringFree-text search over title + abstract.
tagsstring[]Filter to papers tagged with any of these (OR match).
min_scorenumber 0-10Minimum safety score. (default: 5)
limitinteger 1-100Max papers to return. (default: 20)
get_paper

Fetch the full record of a single paper by its directory ID.

Arguments
id *integerPaper ID from the directory.

Prefer raw data?

The full classified dataset is also published as a read-only SQLite snapshot, refreshed nightly. Load it in Datasette, a Jupyter notebook, or any SQL tool — no MCP client required.

Download aisafety.sqlite

What is MCP?

The Model Context Protocol is an open standard for connecting AI assistants to external data sources and tools. A model that supports MCP (like Claude) can call a server's tools mid-conversation — turning queries against this directory into a natural part of chatting with the assistant.