paperdaily MCP server — quickstart

paperdaily-mcp is a stdio Model Context Protocol server that exposes the paperdaily v1 REST API as 24 agent tools. Drop it into Claude Desktop / Claude Code / Cursor / any MCP-compatible runtime and the agent can auto-discover and call paperdaily_get_digest_today, paperdaily_get_paper, paperdaily_ask, etc. directly.

It is a thin proxy: every tool call is forwarded as a typed HTTP request with your bearer key. Accounts, scopes and rate limits are the same ones the website uses — enforced server-side per key, nothing to configure in the client.

Install

From PyPI (recommended):

uv tool install paperdaily-mcp
# or
pipx install paperdaily-mcp

From the public skills repo:

uv tool install \
    --from 'git+https://github.com/alonegg/paperdaily-skills#subdirectory=mcp' \
    paperdaily-mcp

Confirm the binary is reachable:

which paperdaily-mcp

Get your API key

Sign in at https://www.paperdaily.org, then go to 账号设置 → API Keys (https://www.paperdaily.org/account) and issue a key. Pick the scopes you need — read:digest, read:paper cover all read tools, add write:profile for follow/feedback tools and synth:ask for paperdaily_ask. Keep the pd_live_… string — it is shown only once, and the MCP server reads it from the env on launch.

Configure Claude Desktop

Open ~/Library/Application Support/Claude/claude_desktop_config.json (macOS) or %APPDATA%\Claude\claude_desktop_config.json (Windows) and add a paperdaily entry under mcpServers:

{
  "mcpServers": {
    "paperdaily": {
      "command": "paperdaily-mcp",
      "env": {
        "PAPERDAILY_API_KEY": "pd_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx",
        "PAPERDAILY_BASE_URL": "https://www.paperdaily.org"
      }
    }
  }
}

PAPERDAILY_BASE_URL defaults to https://www.paperdaily.org; a self-hosted deployment can point it at its own origin instead.

Restart Claude Desktop. In a new conversation, open the tools list — you should see paperdaily with 24 tools.

Configure Claude Code

claude mcp add paperdaily \
    -e PAPERDAILY_API_KEY=pd_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx \
    -e PAPERDAILY_BASE_URL=https://www.paperdaily.org \
    -- paperdaily-mcp

Verify:

claude mcp list
# paperdaily — stdio — 24 tools

Sanity check from the terminal

The MCP server only relays HTTP, so the fastest end-to-end check is the API itself:

curl -sS -H "Authorization: Bearer pd_live_xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx" \
    https://www.paperdaily.org/api/v1/taxonomy/fields | head -c 400

You should see the top OpenAlex fields with paper counts in the millions. If that works but the agent sees no tools, the problem is on the MCP config side (env not reaching the process, binary not on PATH).

The 24 tools

Mirror the v1 REST surface 1:1. Names follow the paperdaily_<verb>_<noun> convention.

Quotas are enforced server-side per (key tier, scope) — the same table as the REST surface. 429 responses carry Retry-After and X-RateLimit-* headers. Full rate-limit and error-code reference: https://www.paperdaily.org/docs/api/quickstart.html.

Troubleshooting

• PAPERDAILY_API_KEY env var is required — the env var isn't reaching the server. Claude Desktop only honours env: inside the mcpServers block; shell exports won't propagate.
• All tools return HTTP 401 — the key is rejected. Check the prefix (pd_live_…), or list your keys on the account page to confirm it isn't revoked.
• HTTP 403 with missing scope: … — the key wasn't issued with that scope; add the scope on the account page (existing keys can be edited) or issue a new key.
• HTTP 422 on paperdaily_list_papers with q= — q= requires one of topic_id / subfield_id / field_id; add a facet to scope.
• transport error — the binary couldn't reach PAPERDAILY_BASE_URL. Test with the curl sanity check above.