Your config + 6 curated tools in one repo. One-click deploy. Customize freely. Auto-sync upstream.
Обзор
Your config + 6 curated tools in one repo. One-click deploy. Customize freely. Auto-sync upstream. : macOS (Homebrew) and Linux (apt / dnf / pacman). setup.sh auto-detects your platform. 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 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. If you skipped API key configuration during install (e.g., via curl | bash): setup.sh is idempotent — it reads .
README
Why ResearchMate?
| Pain point | Status quo | ResearchMate |
|---|---|---|
| Reconfigure Claude on every new machine | Manual copy, no version control | One command sets up everything |
| Want to use multiple community tools | Installed separately, scattered | 6 tools unified in one repo |
| Community tool gets updated | Fork and manually merge | GitHub Actions auto-sync + Claude review |
| Customized a tool, can’t sync anymore | Locked out of upstream | git subtree – edit freely, still mergeable |
| Dependency setup takes many steps | Read each README step by step | Auto-detect platform, deps, MCP registration |
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:
| Key | Required? | Purpose |
|---|---|---|
ANTHROPIC_API_KEY |
For auto-review | GitHub Actions PR auto-review via Claude |
OPENAI_API_KEY |
For ARIS | Cross-model adversarial review (Codex MCP) |
LLM_API_KEY |
For ARIS | LLM-chat MCP (DeepSeek / Kimi / MiniMax) |
LLM_API_BASE |
With LLM_API_KEY | API base URL for LLM-chat |
MINIMAX_API_KEY |
For ARIS | MiniMax-chat MCP |
HF_TOKEN |
For models | Hugging Face model downloads |
WANDB_API_KEY |
For tracking | Weights & Biases experiment tracking |
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
| Feature | How it works |
|---|---|
| One-click install | curl | bash – auto-detects platform, Claude path, dependencies |
| 6 curated tools | Hand-picked third-party tools in one repo, all pre-configured |
| Auto upstream sync | GitHub Actions runs weekly + Claude Haiku reviews changes + auto-merge |
| Free to customize | Edit any third-party code; git subtree keeps upstream merges clean |
| Priority conflict resolution | Same-name skills auto-resolved: your config > ARIS > academic > scientific |
| Symlink, not copy | Edit repo files, changes take effect instantly – no reinstall needed |
Integrated Tools
Auto-claude-code-research-in-sleep (ARIS) – AI Research Automation Engine
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
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
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
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
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
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
| Feature | Type | Source | Description |
|---|---|---|---|
/idea-discovery |
Skill | ARIS | Literature survey → brainstorm → novelty check → GPU pilot |
/research-pipeline |
Skill | ARIS | End-to-end: idea → implement → review → submit |
/autoresearch |
Skill | config | Karpathy-style autonomous ML experiments |
/autoresearch-loop |
Skill | config | Domain-agnostic autonomous optimization loop |
/implement-paper |
Command | config | Extract algorithms from papers → PyTorch code |
/analyze-paper |
Command | config | Deep analysis of AI papers → structured report |
| Codex / LLM-chat / MiniMax | MCP | ARIS | Cross-model adversarial review |
Paper Writing
| Feature | Type | Source | Description |
|---|---|---|---|
/paper-writing |
Skill | ARIS | Narrative → outline → figures → LaTeX → PDF |
/paper-plan /paper-write /paper-figure /paper-compile |
Skill | ARIS | Individual paper pipeline stages |
/auto-paper-improvement-loop |
Skill | ARIS | Autonomous review → fix → recompile (2 rounds) |
/academic-paper |
Skill | academic-research | 12-agent writing system + bilingual abstracts |
/research-paper-workflow |
Skill | config | Paper structure, LaTeX best practices |
/ai-research-slides |
Skill | config | Paper analysis → academic presentation |
| 19 format templates | Skill | scientific-skills | Nature / Science / NeurIPS / ICLR… |
Literature Review & Peer Review
| Feature | Type | Source | Description |
|---|---|---|---|
/auto-review-loop |
Skill | ARIS | 4-round GPT adversarial review |
/research-review |
Skill | ARIS | Deep critical review via Codex MCP |
/deep-research |
Skill | academic-research | 13-agent PRISMA literature review |
/academic-paper-reviewer |
Skill | academic-research | 7-agent peer review (0-100 scoring) |
/academic-pipeline |
Skill | academic-research | 10-stage orchestrator with integrity checks |
/review-paper |
Command | config | Generate ICLR/ICML/NeurIPS review comments |
Experiment Management & Training
| Feature | Type | Source | Description |
|---|---|---|---|
/run-experiment |
Skill | ARIS | Remote GPU deployment + monitoring |
/train |
Command | config | GPU check → config validation → launch training |
/debug-cuda |
Command | config | CUDA memory / device issue diagnosis |
/ablation |
Command | config | Ablation study design and comparison tables |
/eval-model |
Command | config | Standard benchmark evaluation (FID/MMLU/VQA) |
/benchmark |
Command | config | Run standard ML benchmarks |
/experiment-management |
Skill | config | wandb/tensorboard + hydra config management |
/gpu-optimization |
Skill | config | Mixed precision, gradient accumulation, parallelism |
/pytorch-patterns |
Skill | config | Model architecture, DDP/FSDP, checkpointing |
/dataset-processing |
Skill | config | Large-scale data loading, augmentation, streaming |
Code Review & Quality
| Feature | Type | Source | Description |
|---|---|---|---|
/review-loop |
Command | claude-review-loop | Codex independent review after task completion |
/code-review |
Command | config | Code quality and security review |
/test-coverage |
Command | config | Test coverage analysis (80%+ target) |
/refactor-clean |
Command | config | Dead code identification and safe removal |
| code-reviewer | Agent | config | Code quality review specialist |
| security-reviewer | Agent | config | Security vulnerability detection |
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
| Feature | Type | Description |
|---|---|---|
/sync-upstream |
Claude Code command | Sync third-party tools with upstream from within Claude |
/add-tool |
Claude Code command | Add new third-party tool from within Claude |
/research-mate-update |
Claude Code command | Pull latest changes + re-run setup |
sync-upstream.sh |
Shell script | Manual sync (all or single tool) |
add-tool.sh |
Shell script | Manual add new tool via git subtree |
Statistics
| Type | Count | Sources |
|---|---|---|
| Skills | 230+ | config: 24, ARIS: 27, academic: 4, scientific: 175 |
| Commands | 25 | config: 23, review-loop: 2 |
| Agents | 7 | config |
| MCP Servers | 3 | ARIS: Codex, LLM-chat, MiniMax |
| Rules | 13 | config |
| Contexts | 4 | config |
Dependencies
setup.sh auto-detects your platform and provides install hints.
| Dependency | Required? | macOS | Linux (apt) |
|---|---|---|---|
| Git | Yes | xcode-select --install |
sudo apt install git |
| Node.js | Yes | brew install node |
sudo apt install nodejs npm |
| Claude Code | Recommended | npm i -g @anthropic-ai/claude-code |
same |
| Codex CLI | For ARIS | npm i -g @openai/codex |
same |
| jq | For review-loop | brew install jq |
sudo apt install jq |
| LaTeX | For paper writing | brew install --cask mactex |
sudo apt install texlive-full |
| Python 3 | For literature search | brew install python |
sudo apt install python3 python3-pip |
| uv | For autoresearch | curl -LsSf https://astral.sh/uv/install.sh | sh |
same |
| PyTorch | For autoresearch | pip install torch |
same |
Fedora (
dnf) and Arch (pacman) are also supported –setup.shdetects 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
| Capability | everything-claude-code | dot-claude | starter-kit | ResearchMate |
|---|---|---|---|---|
| Store your own config | ~ | ✅ | ~ | ✅ |
| Integrate multiple third-party tools | ❌ | ❌ | ❌ | ✅ (6 tools) |
| Independent upstream sync per tool | ❌ | ❌ | ~ (one only) | ✅ |
| Local edits still mergeable with upstream | ❌ | N/A | ✅ | ✅ |
| One-click new machine deploy | ✅ | ✅ | ~ | ✅ |
| AI researcher workflows | ❌ | ❌ | ❌ | ✅ |
| Claude auto-reviews sync PRs | ❌ | ❌ | ❌ | ✅ |
| In-Claude maintenance commands | ❌ | ❌ | ❌ | ✅ |
Design Principles
- One-click > Multi-step – Everything configurable in a single command, no manual steps
- Symlink > Copy – Edit repo files, instant effect, no reinstall
- Your config > Third-party – Same-name conflicts always resolve in favor of your config
- Auto > Manual – Upstream sync, conflict detection, dependency checks are all automated
- Backup > Overwrite – Existing config is backed up to
~/.claude/backup/, nothing is lost
License
Acknowledgments
Built on top of these excellent projects:
Рекомендуемые инструменты
Попробуйте другой запрос или уберите фильтр.
Установка
npx skillfish add yxbian23/research-mate