Research & science harnesses for AI agents
33 open-source Research & science harnesses an AI agent can use — MCP servers, SDKs, and adapters. Browse them on Loadbay. An agent can search these over Loadbay's MCP:
claude mcp add --transport http loadbay https://loadbay.xyz/api/mcp
→ Best Research & science harnesses (top picks, ranked)
- scientific-agent-skills — A library of 161 pre-built agent skills covering biology, chemistry, genomics, drug discovery, and clinical research, compatible with Claude Code, Cursor, and Codex. Agents install skills via npx to gain access to 100+ scientific databases and specialized research analysis workflows.
- DeepTutor — Agent-native learning workspace explicitly designed to be operated by other AI agents, connecting tutoring, research, quiz generation, visualization, and mastery practice with JSON output modes for agent integration.
- gpt-researcher — Autonomous deep-research agent that plans and executes multi-step web research on any topic, synthesizing comprehensive reports from multiple LLMs and sources that other agents can act on.
- AI-Scientist — First end-to-end system for fully automated open-ended scientific discovery: generates ideas, runs experiments, and writes papers.
- STORM — Stanford knowledge-curation system that researches a topic via internet search and writes a full-length, cited report.
- Feynman — Open-source AI research agent CLI with a workbench, local and hosted model providers, and installable research skills for Codex.
- paper-qa — High-accuracy RAG framework for answering questions from scientific papers with grounded citations.
- AlphaFold 3 — Official DeepMind inference pipeline for AlphaFold 3, predicting joint structures of proteins, nucleic acids, ligands, and ions.
- AI-Scientist-v2 — Agentic successor that uses progressive tree search to autonomously generate hypotheses, run ML experiments, and write manuscripts.
- AIPOCH Open-Science — Local-first, model-agnostic AI research workbench with scientific agents, Python/R notebooks, data connectors, and reproducible provenance.
- Boltz — Open family of biomolecular interaction models for predicting protein and complex structures and binding affinity, an AlphaFold3-class tool.
- zotero-mcp — MCP server linking a Zotero research library to AI assistants for paper search, summaries, and citation analysis.
- Biomni — General biomedical AI agent that combines LLM reasoning, retrieval-augmented planning, and code execution over a biomedical toolbox.
- arxiv-mcp-server — An MCP server for searching and analyzing arXiv papers, so a research agent can pull and read the literature.
- Hyperresearch — Deep-research harness that turns Claude Code into a multi-step research agent with verified citations and a persistent searchable research vault (~2,300 GitHub stars).
- ESM3 — EvolutionaryScale flagship protein language models (ESM3, ESM C) reasoning jointly over sequence, structure, and function.
- paper-search-mcp — MCP server to search and download academic papers across 20+ sources including arXiv, PubMed, bioRxiv, and Semantic Scholar.
- wisp-science — Open-source, local-first desktop AI research workbench for scientific computing with built-in MCP. Supports Python/R execution and lets agents query scientific databases and run analyses in a sandboxed environment.
- OpenResearch — Local-first workspace that turns Claude Code, Codex, or OpenCode into parallel research agents for literature review, experiments, and autoresearch loops (~1,100 GitHub stars).
- chemcrow-public — LLM chemistry agent that augments models with expert tools for synthesis planning and molecule discovery.
- Denario — A modular multi-agent system that assists with scientific research end to end.
- openresearch-cli — CLI framework that runs parallel AI research agents to review literature, develop hypotheses, run experiments, and produce research artifacts. Multiple agents explore independent directions in isolated git worktrees using Claude Code, Codex, or OpenCode as the underlying model.
- BioAgents — An AI-scientist framework for autonomous deep research in the biological sciences.
- GenoMAS — Multi-agent framework that automates gene-expression analysis workflows end to end.
- CASSIA — Multi-agent LLM framework for automated cell-type annotation of single-cell RNA-seq data.
- BuildArena — Benchmark where LLM agents design, build, and test rockets, cars, and bridges in a physics simulator from text goals.
- agentic-pymol — A lightweight MCP server that exposes PyMOL as a typed tool surface for AI agents, enabling programmatic control of molecular visualization sessions. Returns structured data (coordinates, distances, RMSDs, sequences) rather than plain text for downstream agent reasoning.
- ProteinMCP — Agentic framework for autonomous protein design wrapping structure-prediction and sequence-design tools.
- cactus — LLM agent that leverages cheminformatics tools to answer chemistry questions with informed tool use.
- atlas — Bayesian optimization brain for self-driving laboratories that plans experiments for autonomous discovery.
- AlphaFold-MCP-Server — MCP server exposing the AlphaFold Protein Structure Database for structure-prediction analysis through agent tools.
- semantic-scholar-mcp — MCP server for searching and analyzing academic papers and citation graphs via the Semantic Scholar API.
- MatAgent — Physics-aware multi-agent LLM framework for accelerating materials-science discovery and optimization.