omr

Oh-My-Research (omr)

Research workflows and skills for Claude Code. The plugin's namespace is omr, so every bundled skill is invoked as /omr:<skill-name>.

Install

claude --plugin-dir /path/to/Oh-My-Research

For persistent installation via a marketplace, see the Claude Code plugin docs at https://code.claude.com/docs/en/plugins.

Bundled MCPs

The plugin ships an .mcp.json declaring five MCP servers, loaded automatically when the plugin is enabled:

Server	Transport	Source	Env var
Exa	stdio	exa-mcp-server (npx)	EXA_API_KEY
Tavily	stdio	tavily-mcp (npx)	TAVILY_API_KEY
Brave Search	stdio	@brave/brave-search-mcp-server (npx)	BRAVE_API_KEY
GitHub	http	https://api.githubcopilot.com/mcp/	GITHUB_PERSONAL_ACCESS_TOKEN
Hugging Face	http	https://huggingface.co/mcp	HF_TOKEN

Stdio servers are fetched on first use by npx; HTTP servers are hosted by the vendor and need only a token.

Configuring tokens

Two layers, with shell environment always winning:

1. Shell profile (preferred — required for the HTTP servers) — export in ~/.zshrc / ~/.bashrc:

   export EXA_API_KEY=...
   export TAVILY_API_KEY=...
   export BRAVE_API_KEY=...
   export GITHUB_PERSONAL_ACCESS_TOKEN=...
   export HF_TOKEN=...
   

2. Project-local .env (stdio servers only) — copy .env.example to .env in the directory where you launch Claude Code. The loader shim picks up any keys not already set in your shell. The two HTTP MCPs (github, huggingface) cannot read .env because Claude Code expands their Authorization header from its own process env at startup; for those, use shell env or a tool like direnv that exports before launch.

Token sources:

• Exa https://dashboard.exa.ai/api-keys
• Tavily https://app.tavily.com/home
• Brave https://api.search.brave.com/app/keys
• GitHub PAT https://github.com/settings/personal-access-tokens/new — grant only the scopes you want the MCP to use
• Hugging Face https://huggingface.co/settings/tokens

Missing a token only disables that one server; the others still work. Confirm servers are up via /mcp inside Claude Code — each should show as connected with its tools listed.

Skills

Slash command	Keyword trigger	What it does
/omr:setup	omr-setup	Install, configure, and health-check omr — CLAUDE.md install, MCP token audit, .env scaffolding, optional HPC config.
/omr:doctor	omr-doctor	Read-only diagnostic: plugin integrity, MCP wiring, CLAUDE.md state, token presence, HPC drift, external deps. Points at /omr:setup for fixes.
/omr:literature-review	omr-literature-review	Search → screen → summarize for a research question. KB-first (Zotero, local PDFs) then MCP fan-out (Exa, Tavily, Brave, HF, GitHub) then web. Reproducible screening rubric; re-runs append safely (Search dedups); --from-existing seeds from Zotero/BibTeX.
/omr:sync-overleaf	omr-sync-overleaf	Two-way sync a local paper dir with Overleaf via pyoverleaf (cookie file or browser/keychain auth), driven by .omr/config.yaml. Resolves auth from a pointer (never the secret); dry-run + consent before any push.
/omr:grill-me	omr-grill-me	Reviewer/committee-style grilling that stress-tests a research idea, hypothesis, or plan against your literature corpus + a rubric (novelty, feasibility, methodology, baselines, evaluation, threats). One question at a time; ends with a sharpened claim + ranked risks.

More skills land here as they're added; all follow the /omr:<name> convention.

Layout

Oh-My-Research/
├── .claude-plugin/
│   └── plugin.json
├── .mcp.json           # search MCP server declarations
├── bin/                # plugin-internal scripts (env loader, etc.)
├── skills/             # add <name>/SKILL.md folders here
├── agents/             # optional agent definitions
├── hooks/              # optional hooks.json
└── README.md

Only plugin.json lives inside .claude-plugin/. All component directories (skills, agents, hooks, etc.) sit at the plugin root.

License

MIT