Research Agent Stack

Install preview: npx @ljrwcx99/research-agent-stack init --dry-run Status: 0.1.0-rc.1 Safety marker: dry-run-first

Research Agent Stack is a release-candidate Claude Code configuration package for research workflows. It packages research-focused skills, agents, slash commands, safety hooks, MCP recommendations, and validation assets so a user can preview a complete research workspace before copying anything into a Claude Code configuration directory.

What is it

This project is an evidence-driven, redistributable configuration stack for Claude Code. It is meant to help with literature review, systematic review, citation verification, reproducibility, data and statistical audit, code reproduction, academic writing, peer review, grant writing, and multilingual scholarship.

The stack is distributed as the scoped npm package @ljrwcx99/research-agent-stack and can also be run from a local clone. The CLI is dry-run-first: preview the files, inspect conflicts and warnings, then apply only to an explicit target directory when you decide the plan is safe.

Features

The 0.1.0-rc.1 preview includes:

• 15 research skills in config/skills/.
• 15 specialized research agents in config/agents/.
• 18 slash commands in config/commands/.
• 9 PowerShell safety and workflow hooks in config/hooks/.
• 13 recommended MCP entries and 5 optional MCP entries in config/mcp/.
• 233 scored candidate records in registry/candidates.jsonl.

Quick start

Preview the install plan without writing files:

npx @ljrwcx99/research-agent-stack init --dry-run

Preview a specific target:

npx @ljrwcx99/research-agent-stack init --dry-run --target ~/.claude

Apply only after reviewing the dry-run output and only with an explicit target:

npx @ljrwcx99/research-agent-stack init --target ~/.claude --apply

For a local clone, use the packaged CLI entry directly:

node scripts/cli.js init --dry-run

Safety model

• Dry-run is the default workflow and the published status marker is dry-run-first.
• Apply mode requires an explicit --target path.
• Existing files are not overwritten by default.
• Conflicts are reported before writes; unresolved conflicts abort the apply.
• The installer copies configuration files only. It does not silently edit settings.json, shell profiles, PATH, global npm or pip modules, credentials, or external services.
• Hooks are copied as files, not auto-enabled.
• MCP entries are templates and recommendations. Verify each server, auth requirement, and privacy impact before enabling it.

What gets installed

Apply mode copies this package's configuration tree into the target directory:

skills/       research workflow skills
agents/       specialized research agents
commands/     slash-command Markdown files
hooks/        PowerShell hook scripts
mcp/          MCP recommendation JSON files and templates when present

It does not install API keys, tokens, private data, package-manager globals, Claude Code settings.json, or external MCP servers.

Current status

0.1.0-rc.1 is a release-candidate preview for validating the packaging, installer behavior, documentation, safety model, and selected research configuration. Treat the configuration as useful but still reviewable: candidate MCP templates need local verification, hooks require manual enabling, and apply currently relies on no-overwrite plus abort-on-conflict rather than backup/rollback.

Validation

From a local checkout:

npm test
npm run validate
npm run dry-run
npm run publish-check
npm publish --dry-run

The npm publish dry-run is required before a real release. None of the validation commands should publish externally or modify the current user's Claude Code configuration.

Known limitations

• Apply does not implement backup or rollback in 0.1.0-rc.1; the current safety model is no-overwrite plus abort-on-conflict.
• Hooks are copied but not auto-enabled because hook activation changes runtime behavior and should remain a manual user decision.
• MCP templates and recommendations must be verified before use, especially when they require credentials or send data to external services.
• The package does not configure paid databases, authenticated services, API keys, or private research repositories.

Docs links

• Getting started
• Install guide
• Validation guide
• Publishing guide
• Security model
• Selected configuration
• Evidence report