Paper Scoring Skills

One canonical Agent Skill for using the Paper Scoring MCP server from Codex, Claude Code, OpenClaw, and Hermes. The installer copies files and never relies on symlinks, so it works on Ubuntu and Windows.

Install

pip install "paper-scoring-skills[openai]==0.2.0"
paper-scoring-skills install --client codex --scope user

Use the ollama or all extra instead when appropriate. Supported clients are codex, claude, openclaw, and hermes. For a project installation, run from the repository or OpenClaw workspace root:

paper-scoring-skills install --client claude --scope project

Hermes project installs use .hermes/skills and require an explicit trust decision from inside the repository:

paper-scoring-skills install --client hermes --scope project
hermes skills trust

Connect the MCP server

The Skill deliberately does not embed credentials or a fixed remote URL. Register the installed local server over STDIO. The following examples show runtime dotenv injection; replace the sample absolute path with the appropriate POSIX or Windows path. Omit the final two server arguments when no dotenv file is needed.

Codex:

codex mcp add paper-scoring -- paper-scoring-mcp --transport stdio --env-file /absolute/path/runtime.env

Claude Code:

claude mcp add --transport stdio --scope user paper-scoring -- paper-scoring-mcp --transport stdio --env-file /absolute/path/runtime.env

OpenClaw:

openclaw mcp add paper-scoring --command paper-scoring-mcp --arg=--transport --arg=stdio --arg=--env-file --arg=/absolute/path/runtime.env
openclaw mcp doctor paper-scoring --probe

For Hermes, add this entry under mcp_servers in ~/.hermes/config.yaml, then run /reload-mcp:

paper_scoring:
  command: paper-scoring-mcp
  args: [--transport, stdio, --env-file, /absolute/path/runtime.env]

The installed Skill contains no credentials. Configure the MCP server through its process environment or explicit --env-file option.

Discuss a daily digest

Version 0.2 also installs paper-scoring-digest. In an OpenClaw Slack thread, the Skill maps a reply such as 3, 7: compare the methods to the exact stored run and ranks. It reads retained PDFs for page-level discussion and does not silently rescore the papers. PDFs expire after the digest service's configured retention period; metadata, scores, and arXiv links remain available.