Skip to content

Installation

  • Claude Code or Claude Cowork on a paid plan.
  • LaTeX compilation: any TeX installation (see LaTeX engine below).
  • Word output: node (for templates/word/build-docx.js, built on the docx library).
  • The evidence kernel: optional (see The evidence kernel below).

To compile-check LaTeX you need a TeX installation. Any of these works, and if you already have a TeX distribution you do not need to install anything else.

Tectonic (recommended). A single binary that fetches the packages it needs on demand, which keeps builds small and reproducible. It is also what CI uses.

Terminal window
brew install tectonic

A full TeX distribution. TeX Live (Linux, Windows, macOS), MiKTeX (Windows first, also macOS and Linux), and MacTeX (macOS) all ship latexmk, which the plugin drives automatically, including the rerun and bibliography passes.

The compile scripts autodetect whichever engine you have. To override the choice, set the LATEX_ENGINE environment variable or pass --engine:

Terminal window
LATEX_ENGINE=latexmk python scripts/latex-compile.py manuscript/main.tex
python scripts/latex-compile.py --engine xelatex manuscript/main.tex
  1. Add the marketplace:

    Terminal window
    /plugin marketplace add sokolmarek/researcher
  2. Install the plugin:

    Terminal window
    /plugin install researcher@researcher-marketplace

Since 0.3.0 the plugin ships core/, a deterministic Python package that queries eight scholarly indexes directly and verifies references against them. It powers multi-index citation verification, retraction and correction checks, open-access resolution, and the provenance ledger.

You do not have to install it. Without it, the plugin works exactly as it did in 0.2.0: the scripts fall back to their standard-library behavior, and nothing hard-fails. Install it when you want citation verification that consults more than one index.

The recommended path is uv. There is no install step at all: uv provisions the environment from core/pyproject.toml the first time a skill calls the kernel.

Terminal window
curl -LsSf https://astral.sh/uv/install.sh | sh

Without uv, install it with pip from a clone:

Terminal window
pip install -e core/ # base runtime: httpx, rapidfuzz, platformdirs
pip install -e "core/[fulltext]" # plus text extraction from open-access PDFs

Check it works:

Terminal window
uv run --project core python -m researcher_core --help

The kernel is keyless. Setting a contact address puts your requests in each API’s polite pool, which raises your rate limits and costs nothing:

Terminal window
export OPENALEX_MAILTO="you@example.edu"
export CROSSREF_MAILTO="you@example.edu"
export UNPAYWALL_EMAIL="you@example.edu"

What it does not do: semantic search over paper contents (embeddings, vector stores, reranking). That layer is deliberately deferred. See the Roadmap, and evals/BENCHMARKS.md in the repository for what the kernel measurably does and does not get right.

Some skills are richer when a connector is available. Configure the ones you use:

ConnectorHowNotes
Scite (Smart Citations)MCP serverScite account
Zotero (library sync)uvx zotero-mcpZOTERO_API_KEY, library id
Paper search (arXiv, PubMed, S2, and more)uvx paper-search-mcpkeyless, free-first

Google Scholar and Mendeley are documentation-only (no stable free API); the relevant skills fall back to web search and manual export.

Ask Claude something small and see the right skill fire:

Find recent papers on self-supervised learning for ECG classification

If a literature search kicks off, you are ready.

For terminal and IDE commits (outside Claude Code), run python scripts/install-git-hooks.py once per manuscript repository so the citation guard also covers terminal and IDE commits.

Next: the Quick Start.