ARH — AI Research Helper

ARH (AI Research Helper) is a CLI-first research assistant agent for AI researchers. It turns everyday research actions — reading papers, searching code, running experiments, writing summaries — into a runtime with tool calling, context management, approval-gated writes, and persistent sessions.

All state lives locally under ~/.arh/; your conversations, memories, and API keys never leave your machine.

• Repository: github.com/LiRunGuo/Arhelper
• Language / Stack: Python 3.9+, Pydantic, FastAPI, SQLAlchemy, OpenAI / Anthropic / Ollama SDKs
• License: MIT
• Status: 0.2.0-alpha (actively developed)

Highlights

• Plan / Execute safety model — by default only read-only tools are exposed; write-class actions must first produce a plan and pass a user approval gate, so the agent never silently modifies your files.
• Three-layer memory — regex + LLM automatic extraction of user preferences and research directions; BM25 cross-session recall; end-of-session LLM reflection summary, auto-injected next time.
• Skill self-learning — successful complex tasks can be saved (after approval) as reusable Skills, callable via /skill or by the LLM.
• Multi-LLM + fallback — switch between OpenAI, Anthropic, and Ollama with one command; automatic cooldown-based failover to a backup model.
• Automatic context compression — when tokens exceed budget, an LLM summarises older turns while preserving the system prompt, recent rounds, and tool-call pairs.
• Checkpoint & resume — snapshot and --resume any session; no lost context on interruption.
• Multi-channel gateway — Hub-Spoke architecture; CLI and Telegram shipped, Feishu and WeChat in progress; HTTP endpoints exposed via FastAPI.
• Research-native tools — arXiv / HuggingFace / GitHub / OpenReview crawlers, deep paper interpretation, topic surveys, trend analysis, daily digest — all built in.

Quick install

git clone https://github.com/LiRunGuo/Arhelper.git
cd Arhelper
pip install -e .
arh doctor

Or run via Docker Compose (docker compose up).