refhub

// structured_research_memory for humans, tools, and agents

refhub is a modern reference management platform for organizing academic publications, building citation networks, and sharing research collections.

it is built for the command-line generation: browser-first, keyboard-friendly, dark by default, structured enough for agents and scripts, and blunt about the difference between user workflows and automation workflows.

refhub is now a multi-repo system with a clear split between product, api, cli, extension, and agent integration work.

• refhub.io is the frontend product: vaults, papers, profiles, settings, help, and api-key management
• .netlify is the backend/api layer: /api/v1, scoped keys, semantic scholar routes, pdf/item routes, import/export surfaces
• refhub-cli and refhub-skill are the automation layer for humans, scripts, and agents
• refhub-extensions handles browser capture
• refhub-paper-drafter is a specialized agent skill built on top of refhub-skill/refhub-cli, for drafting HCI and visualization research papers from a vault
• agent skills self-host their own Claude Code and Codex plugin marketplaces — no separate registry repo

near-term work is focused on cleaner vault workflows, stronger permission semantics, better import/export/search, and safer agentic operations.

refhub is not a prettier paper list. the target is a research memory system with explicit structure:

• vault-based publication management
• public and private research collections
• tags, metadata, relations, profiles, and settings
• pdf and google drive attachment workflows
• versioned api access for trusted automation
• cli, browser extension, and agent skill surfaces

in plain english: usable in the browser, programmable without regret.

each agent skill repo is self-contained — no shared registry. install refhub-skill for general vault/api operation, refhub-paper-drafter for the paper-drafting workflow (works standalone, or layered on top of refhub-skill).

claude code

claude plugin marketplace add https://github.com/refhub-io/refhub-skill
claude plugin install refhub-skill@refhub-skill

claude plugin marketplace add https://github.com/refhub-io/refhub-paper-drafter
claude plugin install refhub-paper-drafter@refhub-paper-drafter

codex

add to your Codex plugin marketplace configuration (~/.agents/plugins/marketplace.json or project-local):

{
  "name": "refhub-skill",
  "source": { "source": "github", "repo": "refhub-io/refhub-skill" },
  "policy": { "installation": "AVAILABLE", "authentication": "ON_INSTALL" },
  "category": "Productivity"
}

each repo also ships its own .agents/plugins/marketplace.json, so cloning locally and pointing Codex at the clone works too — see each repo's README.

gemini cli · opencode

# gemini
mkdir -p ~/.gemini/skills/refhub-skill && curl -o ~/.gemini/skills/refhub-skill/SKILL.md \
  https://raw.githubusercontent.com/refhub-io/refhub-skill/main/skills/refhub-skill/SKILL.md

# opencode
mkdir -p ~/.config/opencode/skills/refhub-skill && curl -o ~/.config/opencode/skills/refhub-skill/SKILL.md \
  https://raw.githubusercontent.com/refhub-io/refhub-skill/main/skills/refhub-skill/SKILL.md

cursor · windsurf · pi · other generic harnesses

curl -O https://raw.githubusercontent.com/refhub-io/refhub-skill/main/AGENTS.md

each skill repo's AGENTS.md is a full standalone copy of its skill instructions — no plugin system required.

if a concept matters, it should exist explicitly in the model and api.

tags, relations, permissions, vault roles, scopes, and item metadata should be real things — not hidden conventions.

agents and scripts should get narrow, inspectable access:

• scoped api keys
• vault restrictions
• honest errors
• auditability
• eventually dry-runs, safer bulk operations, and sync surfaces

normal user auth is for people. api keys are for runtime automation. account setup and connected services stay in the web app.

refhub is for researchers and builders who are comfortable with dense interfaces, keyboards, and actual structure.

• dark by default
• keyboard-first
• monospace where data matters
• // as the signature pattern
• low on marketing glaze
• closer to a good terminal tool than a lifestyle app

keep the split clean:

• frontend concerns belong in refhub.io
• backend/api concerns belong in .netlify
• cli ergonomics belong in refhub-cli
• extension concerns belong in refhub-extensions
• integration abstractions should follow the real backend, not outrun it

good contributions improve data structure, permission clarity, workflow efficiency, integration safety, or research usability.