PantheonOS

An evolvable, distributed multi-agent framework and testbed for data science, with a primary focus on end-to-end single-cell biology analysis, developed by Stanford Qiu Lab.

Overview#

PantheonOS is developed by Stanford Qiu Lab, positioned as a "data science operating system." It features a four-layer pyramid architecture (APPLICATION → INTERFACE → AGENT → LLM) built on NATS messaging for cross-machine distributed deployment.

Core Capabilities#

Agentic Code Evolution#

Pantheon-Evolve: Genetic algorithm-driven autonomous improvement of agent algorithms and code, achieving super-human performance on batch correction and RL-enhanced gene panel selection (specific benchmarks pending paper confirmation)
Code analysis powered by tree-sitter (Python / JavaScript parsing)

Multi-Agent Orchestration#

Team patterns: PantheonTeam, Sequential, Swarm, Mixture-of-Agents (MoA), AgentAsTool

Distributed Execution#

Scalable, fault-tolerant deployment on NATS messaging with cross-machine support and WebSocket configuration
MCP protocol integration (fastmcp>=2.11.0)

Domain Analysis#

Single-cell & Spatial Transcriptomics: Preprocessing, clustering, integration, trajectory inference, spatial mapping, SCENIC gene regulatory networks
Multi-omics Integration: MOFA, GLUE, SIMBA, TOSICA for scRNA + scATAC integration
Protein Structure & Drug Design: AlphaFold2 prediction, RFdiffusion generation, BindCraft binder design, ipSAE ranking, drug target validation
TCGA Pan-Cancer Analysis: Bulk RNA-seq preprocessing, survival analysis
Developmental Biology & Disease Mechanisms: Embryonic axis asymmetric paracrine signaling, congenital heart disease molecular mechanisms, virtual cell predictive modeling

Knowledge & Retrieval#

Built-in LanceDB vector database, optional LlamaIndex + Qdrant integration (knowledge extra)
Web crawling & search: crawl4ai + DuckDuckGo

Interaction & Access#

Multi-language mixing: Python, R, Julia + natural language in CLI
Multi-channel gateways: Slack, Telegram, Discord, Lark
Jupyter Notebook execution: Built-in jupyter-client / ipykernel
Desktop apps (macOS / Windows / Linux) and online Web App available

Community Ecosystem#

Pantheon Store: Community marketplace with 1,000+ curated agents, teams, and skills (genomics, pharmacology, medicine, bioinformatics, etc.)

Installation & Deployment#

pip install pantheon-agents

Recommended via uv:

git clone https://github.com/aristoteleo/PantheonOS.git
cd PantheonOS
uv sync

Optional extras:

uv sync --extra knowledge   # RAG/vector search
uv sync --extra claw        # Mobile gateway
uv sync --extra r           # R language support

Docker deployment:

docker pull nanguage/pantheon-agents:latest
docker run -it --rm -e PANTHEON_MODE=standalone -v $(pwd)/workspace:/workspace -p 8080:8080 nanguage/pantheon-agents:latest

Quick Start#

Install via any method above
Configure LLM API Key (e.g., OPENAI_API_KEY); supports OpenAI, Anthropic, Google Genai backends
Run pantheon cli for interactive mode, or pantheon ui --auto-start-nats --auto-ui for multi-agent chatroom UI
Or access the online Web App directly

Privacy#

Fully local deployment supported for sensitive biological data control

Pending Confirmation#

Paper page exists but title, journal, DOI not fully loaded
License discrepancy: pyproject.toml declares MIT, GitHub LICENSE file states BSD-2-Clause (LICENSE file takes precedence)
Specific benchmark numbers for "super-human performance" require paper reference
PyPI URL inferred from package name, not verified