yxbian23/research-mate

---

Why ResearchMate?

Architecture

 ResearchMate
 ├── config/                          Your Claude Code config
 │   ├── 7 agents                     (architect, code-reviewer, planner, ...)
 │   ├── 13 rules                     (gpu-safety, ml-coding-style, security, ...)
 │   ├── 25 commands                  (/train, /plan, /review-paper, ...)
 │   ├── 24 skills                    (pytorch-patterns, experiment-management, ...)
 │   ├── 4 contexts                   (research, training, review, dev)
 │   └── settings.json.template
 │
 ├── third-party/                     6 curated tools (git subtree, auto-sync)
 │   ├── Auto-claude-code-research-in-sleep (ARIS)/  27 skills, 3 MCP servers — AI research automation
 │   ├── autoresearch/                program.md + train.py — Autonomous ML iteration on single GPU
 │   ├── pi-autoresearch/             1 skill, 1 extension — Domain-agnostic autonomous optimization loop
 │   ├── claude-review-loop/          2 commands, 1 stop hook — Automated Codex code review
 │   ├── academic-research-skills/    4 skills (13+12+7-agent systems) — Full academic workflow
 │   └── claude-scientific-skills/    175 scientific data source skills — Scientific database access
 │
 ├── install.sh          curl | bash one-liner
 ├── setup.sh            Local setup (idempotent)
 ├── sync-upstream.sh    Manual upstream sync
 └── add-tool.sh         Add new third-party tool

Quick Start

Supported platforms: macOS (Homebrew) and Linux (apt / dnf / pacman). setup.sh auto-detects your platform.

One-line install

curl -fsSL https://raw.githubusercontent.com/yxbian23/research-mate/main/install.sh | bash

Custom binary paths (common on Linux)

curl -fsSL https://raw.githubusercontent.com/yxbian23/research-mate/main/install.sh | bash -s -- \
  --claude-path /path/to/claude \
  --codex-path /path/to/codex

Or clone manually

git clone https://github.com/yxbian23/research-mate.git ~/.research-mate
cd ~/.research-mate && ./setup.sh

The setup script automatically:

• Checks and reports all dependency status
• Backs up existing ~/.claude/ config (nothing is lost)
• Symlinks config files to ~/.claude/
• Installs all skills with conflict detection and priority resolution
• Registers MCP servers (Codex, LLM-chat, MiniMax)
• Installs Claude Code plugins (document-skills, huggingface-skills)
• Prompts for API key configuration
• Works with any combination: no Claude/Codex, either one, or both installed

Configuration

API Keys

Edit ~/.research-mate/.env to configure API keys:

When running via curl | bash, the API key prompt is skipped automatically. Configure keys afterwards as described below.

Post-Install Setup

If you skipped API key configuration during install (e.g., via curl | bash):

# 1. Edit .env and add your keys
vim ~/.research-mate/.env

# 2. Re-run setup to register MCP servers
cd ~/.research-mate && ./setup.sh

setup.sh is idempotent — it reads .env, detects newly added keys, and registers the corresponding MCP servers (LLM-chat, MiniMax) without affecting existing config.

Alternatively, register MCP servers manually:

# LLM-chat MCP (requires LLM_API_KEY in .env)
claude mcp add llm-chat -s user -- python3 ~/.research-mate/third-party/aris/mcp-servers/llm-chat/server.py

# MiniMax MCP (requires MINIMAX_API_KEY in .env)
claude mcp add minimax-chat -s user -- python3 ~/.research-mate/third-party/aris/mcp-servers/minimax-chat/server.py

Key Features

Integrated Tools

Auto-claude-code-research-in-sleep (ARIS) – AI Research Automation Engine

wanshuiyin/Auto-claude-code-research-in-sleep

27 skills, 3 MCP servers (Codex, LLM-chat, MiniMax). Highlights: cross-model adversarial review (Claude + GPT), idea discovery pipeline, 4-round auto review loop, end-to-end research pipeline.

autoresearch – Autonomous ML Experiments

karpathy/autoresearch

Andrej Karpathy’s framework: give an AI agent a single GPU, let it modify code, train, evaluate, keep-or-discard, repeat. Single-file constraint keeps scope auditable.

pi-autoresearch – Domain-Agnostic Optimization Loop

davebcn87/pi-autoresearch

Not limited to ML – optimize any quantifiable metric (build speed, bundle size, test performance). Persistent state across sessions via autoresearch.jsonl.

academic-research-skills – Academic Workflow System

Imbad0202/academic-research-skills

4 skills: deep-research (13-agent PRISMA literature review), academic-paper (12-agent writing system), paper-reviewer (7-agent peer review), academic-pipeline (10-stage orchestrator).

claude-review-loop – Automated Code Review

hamelsmu/claude-review-loop

2 commands, 1 stop hook. Triggers Codex for independent review after development. Provides third-party perspective on code quality with iterative feedback loops.

claude-scientific-skills – Scientific Data Access

K-Dense-AI/claude-scientific-skills

175 production-ready skills accessing 250+ data sources (PubMed, ChEMBL, UniProt, COSMIC, SEC EDGAR, and more). 19 journal/conference paper format templates.

Feature Catalog

Research Automation

Paper Writing

Literature Review & Peer Review

Experiment Management & Training

Code Review & Quality

Scientific Data Access

175 production-ready skills covering 250+ scientific databases and APIs (PubMed, ChEMBL, UniProt, COSMIC, SEC EDGAR, and more). 19 journal/conference paper format templates.

Office Documents

Skills for PDF, DOCX, PPTX, and XLSX creation, editing, and analysis.

Maintenance

Statistics

Dependencies

setup.sh auto-detects your platform and provides install hints.

Fedora (dnf) and Arch (pacman) are also supported – setup.sh detects them automatically.

Usage Guide

Edit your config

# Edit files in config/ -- symlinks make changes take effect immediately
vim ~/.research-mate/config/rules/my-new-rule.md
git add . && git commit -m "feat: add new rule" && git push

Customize third-party tools

# Edit directly -- changes are instant, and upstream merges still work
vim ~/.research-mate/third-party/aris/skills/auto-review-loop/SKILL.md
git add . && git commit -m "feat: customize ARIS review loop" && git push

Sync upstream updates

# In Claude Code:
/sync-upstream

# Or via shell:
./sync-upstream.sh            # sync all tools
./sync-upstream.sh aris       # sync one tool

# Automatic: GitHub Actions checks weekly, Claude Haiku reviews, auto-merge

Add a new tool

# In Claude Code:
/add-tool

# Or via shell:
./add-tool.sh   [branch]

Update ResearchMate itself

# In Claude Code:
/research-mate-update

# Or via shell:
cd ~/.research-mate && git pull && ./setup.sh

Comparison

Design Principles

1. One-click > Multi-step – Everything configurable in a single command, no manual steps
2. Symlink > Copy – Edit repo files, instant effect, no reinstall
3. Your config > Third-party – Same-name conflicts always resolve in favor of your config
4. Auto > Manual – Upstream sync, conflict detection, dependency checks are all automated
5. Backup > Overwrite – Existing config is backed up to ~/.claude/backup/, nothing is lost

License

MIT

Acknowledgments

Built on top of these excellent projects:

• autoresearch by @karpathy
• pi-autoresearch by @davebcn87
• Auto-claude-code-research-in-sleep (ARIS) by @wanshuiyin
• claude-review-loop by @hamelsmu
• academic-research-skills by @Imbad0202
• claude-scientific-skills by @K-Dense-AI