A curated list of 150+ tools and libraries for building self-driving laboratories and autonomous research platforms.
From GPU compute to lab hardware, from multi-agent orchestration to experiment verification — everything you need to build systems that do science autonomously.
Maintained by Scivity Labs
| Category | Description |
|---|---|
| 🧪 Experiment Orchestration & Workflow | Schedulers, DAGs, and workflow engines that drive long-running scientific pipelines. |
| 🤖 Multi-Agent Frameworks | Libraries for building coordinated agent systems that plan, act, and reason. |
| 🔬 Self-Driving / Autonomous Labs | Reference projects and companies operating real autonomous laboratories. |
| ⚙️ Hardware & Lab Automation | Drivers, protocols, and robotics frameworks that bridge code to physical instruments. |
| 🖥️ GPU Compute Platforms | On-demand and reserved GPU infrastructure for training and inference. |
| 📊 ML Experiment Tracking | Metrics, artifact, and hyperparameter tracking for ML workflows. |
| 🔍 Verification & Reproducibility | Data versioning, validation, and environment-pinning tools. |
| 📚 Scientific Knowledge Management | Literature APIs, reference managers, and citation graph tools. |
| 📄 Scientific Data Extraction | Parsers for PDFs, tables, equations, and scientific figures. |
| 🧠 LLM for Science | Models and agent systems purpose-built for scientific discovery. |
| 🧬 Scientific Simulation & Modeling | Physics-based simulators used as in-silico filters in autonomous-science loops. |
| 🎯 Bayesian Optimization & Active Learning | Decision core of closed-loop autonomous experimentation — picks the next experiment under a budget. |
| 🔗 RL for Scientific Discovery | Reinforcement learning libraries applicable to scientific search problems. |
| 🗄️ Vector Databases & Embeddings | Stores and models for semantic retrieval over scientific corpora. |
| 📡 Data Pipeline & Event Systems | Message brokers and streaming systems for lab and compute events. |
| 🛡️ Safety & Guardrails for Autonomous Systems | Policy enforcement, prompt-injection defense, and output validation. |
| 📈 Monitoring & Observability | Metrics, logs, traces, and LLM-specific observability. |
Orchestrators schedule long-running, failure-prone scientific workflows — think multi-day simulations, instrument sweeps, and data ingestion pipelines. Good ones handle retries, artifact passing, and dynamic task graphs without forcing you into a rigid DSL. Pick one that matches your infra: Kubernetes-native, Python-native, or bioinformatics-specialized.
| Library | Description | Link |
|---|---|---|
| Prefect | Python workflow orchestrator with dynamic task graphs, typed results, and work-pool-based deployment. | github.com/PrefectHQ/prefect |
| Apache Airflow | DAG-based scheduler with a large operator ecosystem, used widely for batch data pipelines. | github.com/apache/airflow |
| Dagster | Asset-oriented orchestrator with typed inputs and outputs, software-defined assets, and testable pipelines. | github.com/dagster-io/dagster |
| Nextflow | Dataflow workflow language widely used in bioinformatics, with first-class HPC and container support. | github.com/nextflow-io/nextflow |
| Snakemake | Python-like rules engine for reproducible scientific workflows, common in genomics. | github.com/snakemake/snakemake |
| Metaflow | Netflix's human-centric framework for data science workflows with versioning and AWS/K8s backends. | github.com/Netflix/metaflow |
| Argo Workflows | Kubernetes-native workflow engine where each step runs in its own container. | github.com/argoproj/argo-workflows |
| Luigi | Spotify's Python module for building complex batch pipelines with dependency resolution. | github.com/spotify/luigi |
| Kedro | Python framework for reproducible, modular data science pipelines with a catalog abstraction. | github.com/kedro-org/kedro |
| Flyte | Kubernetes-native orchestrator for typed, reproducible ML and data pipelines. | github.com/flyteorg/flyte |
| Temporal | Durable execution platform for long-running workflows with retries and versioning as primitives. | github.com/temporalio/temporal |
| Kubeflow Pipelines | ML workflow system for portable, scalable pipelines on Kubernetes. | github.com/kubeflow/pipelines |
Multi-agent frameworks coordinate LLM-driven workers that plan, call tools, and hand off tasks. In an autonomous-science context they sit between the orchestrator and the lab, turning a research goal into concrete experiments. The tradeoff is control vs. flexibility — graph-based frameworks are easier to debug, role-based ones move faster.
| Library | Description | Link |
|---|---|---|
| LangGraph | Graph-based agent runtime from LangChain with checkpoints, streaming, and human-in-the-loop primitives. | github.com/langchain-ai/langgraph |
| CrewAI | Role-based multi-agent framework that assembles specialized agents into crews with shared tasks. | github.com/crewAIInc/crewAI |
| AutoGen | Microsoft's framework for multi-agent conversations with pluggable model backends and code execution. | github.com/microsoft/autogen |
| Swarms | Multi-agent orchestration framework with hierarchical, concurrent, and mixture-of-agents structures. | github.com/kyegomez/swarms |
| CAMEL | Research-oriented multi-agent framework focused on role-playing agents and agent-scaling studies. | github.com/camel-ai/camel |
| Agno | Stateless Python agent runtime with memory, knowledge, and a focus on low-latency instantiation. | github.com/agno-agi/agno |
| smolagents | Hugging Face's minimal agent library centered on code-writing agents. | github.com/huggingface/smolagents |
| PydanticAI | Typed agent framework that uses Pydantic schemas for inputs, outputs, and tool calls. | github.com/pydantic/pydantic-ai |
| OpenAI Agents SDK | OpenAI's Python SDK for building agents with handoffs, guardrails, and tracing. | github.com/openai/openai-agents-python |
| Claude Agent SDK | Anthropic's SDK for building agents on top of Claude with hooks, subagents, and custom tools. | github.com/anthropics/claude-agent-sdk-python |
| BeeAI Framework | IBM-backed framework for building agents in Python and TypeScript with MCP and A2A support. | github.com/i-am-bee/beeai-framework |
| Atomic Agents | Lightweight framework built around typed input/output schemas and composable atomic agents. | github.com/BrainBlend-AI/atomic-agents |
| Langroid | Multi-agent Python framework with first-class tools, vector stores, and message-provenance logging. | github.com/langroid/langroid |
These are operational self-driving laboratories and companies — not libraries, but reference points for what a working autonomous lab looks like. Some publish papers, some ship APIs, a few run cloud labs that anyone can submit experiments to. Study their architectures when designing your own.
| Project | Description | Link |
|---|---|---|
| A-Lab (LBNL / Berkeley) | Autonomous lab for inorganic solid-state synthesis combining robotics, ML, and literature data from the Ceder group. | ceder.berkeley.edu |
| Emerald Cloud Lab | Commercial cloud lab in Austin with 200+ instruments controlled remotely via a Wolfram-based command language. | emeraldcloudlab.com |
| Strateos | Automation-as-a-service cloud lab for drug discovery with programmatic APIs and hybrid on-prem deployments. | strateos.com |
| Chemify | Glasgow-based Chemputation facility that compiles digital code into physical organic-chemistry syntheses. | chemify.io |
| Arctoris | Oxford-based fully automated drug discovery platform running assay cascades on the Ulysses robotic system. | arctoris.com |
| LabGenius | London-based closed-loop platform for therapeutic antibody discovery built around the EVA robot. | labgeniustx.com |
| Kebotix | Boston-area self-driving lab for materials discovery combining generative models with robotic synthesis. | kebotix.com |
| Atinary | Bayesian-optimization and SDLabs software for closed-loop R&D, integrable with external automation. | atinary.com |
| PNNL Autonomous Science | DOE program applying autonomous-lab methods across chemistry, biology, and energy storage at PNNL. | pnnl.gov/autonomous-science |
| Lila Sciences | Flagship Pioneering company building closed-loop AI Science Factories across life, chemical, and materials science. | lila.ai |
| Argonne Polybot | Autonomous robotic platform for electronic-polymer discovery at Argonne's Center for Nanoscale Materials. | anl.gov polybot |
Everything physical bolts onto this layer — pipetting robots, plate readers, mass specs, flow reactors. The libraries here speak protocols (OPC UA, MQTT, VISA, SiLA2) or wrap vendor SDKs. Expect to write glue code; there is no universal driver.
| Library | Description | Link |
|---|---|---|
| opcua-asyncio | Asyncio-based OPC UA client and server for Python, common in industrial and lab PLC integration. | github.com/FreeOpcUa/opcua-asyncio |
| PyVISA | Python bindings to the VISA standard for controlling test and measurement instruments over GPIB, USB, or serial. | github.com/pyvisa/pyvisa |
| paho-mqtt (Python) | Eclipse Paho MQTT client for Python, used for lightweight pub/sub between lab devices and controllers. | github.com/eclipse/paho.mqtt.python |
| Eclipse Mosquitto | Open-source MQTT broker often deployed as the message bus inside automated labs. | github.com/eclipse/mosquitto |
| LabVIEW | National Instruments' graphical programming environment for instrument control and DAQ. | ni.com/labview |
| Opentrons | Python API and firmware for the OT-2 and Flex liquid-handling robots. | github.com/Opentrons/opentrons |
| PyLabRobot | Hardware-agnostic Python SDK for liquid handlers, plate readers, heater-shakers, and scales. | github.com/PyLabRobot/pylabrobot |
| SiLA 2 Python | Python implementation of the SiLA 2 gRPC-based lab instrument interoperability standard. | gitlab.com/SiLA2/sila_python |
| Labman Automation | Commercial systems integrator that built Berkeley's A-Lab synthesis robot and similar custom platforms. | labmanautomation.com |
| ROS 2 | Robot Operating System with DDS-based middleware for real-time control of robotic cells and mobile platforms. | github.com/ros2 |
| Bluesky | Python experiment-control framework for synchrotrons and scattering beamlines, used across NSLS-II. | github.com/bluesky/bluesky |
On-demand GPU platforms spare you from standing up Kubernetes just to run a fine-tune or a large batch inference job. They differ on cold-start latency, GPU availability (H100, B200, MI300), and whether they expose raw VMs or serverless functions. For autonomous-science pipelines, the serverless models integrate cleanly into workflow engines.
| Platform | Description | Link |
|---|---|---|
| RunPod | Community and secure GPU cloud with on-demand pods and serverless endpoints that scale to zero. | runpod.io |
| Modal | Serverless Python platform that provisions GPU containers from decorators in under a second. | modal.com |
| Lambda | GPU cloud focused on H100 and Blackwell clusters with pre-installed ML frameworks and bare metal. | lambda.ai |
| Vast.ai | Marketplace for renting community GPUs at spot-style prices. | vast.ai |
| CoreWeave | Hyperscale NVIDIA-specialized cloud offering H100/B200 clusters and high-throughput networking. | coreweave.com |
| Paperspace (DigitalOcean) | GPU notebooks and droplets folded into DigitalOcean's AI platform. | paperspace.com |
| Lightning AI | Studio-based cloud workspaces on H100/H200 with persistent environments and multi-GPU training. | lightning.ai |
| Anyscale | Managed Ray platform for distributed training, batch inference, and serving on GPU clusters. | anyscale.com |
| Together AI | GPU cloud and inference API focused on open models, fine-tuning, and low-latency serving. | together.ai |
| Replicate | Hosted inference for open-source models with a simple HTTP API and per-second billing. | replicate.com |
Tracking tools log metrics, parameters, artifacts, and code versions so experiments are reproducible and comparable. For autonomous science, the key feature is headless logging from long-running agents — not just interactive notebooks. Many of these integrate with orchestrators above.
| Library | Description | Link |
|---|---|---|
| MLflow | Open-source tracking, model registry, and deployment framework with a large integration ecosystem. | github.com/mlflow/mlflow |
| Weights & Biases | Hosted tracking, artifact, and sweep service with deep SDK integration across frameworks. | github.com/wandb/wandb |
| Neptune.ai | Tracker built for foundation-model training with per-layer metrics and long-run logging. | github.com/neptune-ai/neptune-client |
| Comet | Experiment tracking and LLM evaluation platform with open-source Opik for LLM observability. | comet.com |
| Aim | Self-hosted open-source tracker with a fast UI for comparing thousands of runs. | github.com/aimhubio/aim |
| ClearML | End-to-end platform spanning experiment tracking, data management, agents, and serving. | github.com/clearml/clearml |
| Sacred | Lightweight Python library for configuring, organizing, and logging reproducible experiments. | github.com/IDSIA/sacred |
| DVC | Git-based data and model versioning with experiment tracking and pipeline DAGs. | github.com/iterative/dvc |
| TensorBoard | TensorFlow's local metric and graph visualization tool, also used with PyTorch and JAX. | github.com/tensorflow/tensorboard |
| Optuna | Hyperparameter optimization framework with pruning, distributed trials, and tracker integrations. | github.com/optuna/optuna |
Autonomous systems generate more data, configurations, and artifacts than humans can audit. These tools version datasets, validate schemas, check distributions, and pin environments so a run from six months ago still reproduces. Pair a data-versioning tool with a validation framework; each alone is half the story.
| Library | Description | Link |
|---|---|---|
| Pachyderm | Kubernetes-native pipelines with immutable data versioning and lineage across pipeline stages. | github.com/pachyderm/pachyderm |
| lakeFS | Git-like branching and time travel over object stores for data-lake version control. | github.com/treeverse/lakeFS |
| Great Expectations | Declarative data-quality framework with expectations, profiling, and automated docs. | github.com/great-expectations/great_expectations |
| Evidently AI | Library for data and ML model validation, drift detection, and monitoring reports. | github.com/evidentlyai/evidently |
| Deepchecks | Open-source testing for data and ML models covering integrity, drift, and performance. | github.com/deepchecks/deepchecks |
| pytest | Python testing framework with fixtures, parametrization, and a plugin ecosystem. | github.com/pytest-dev/pytest |
| Hypothesis | Property-based testing for Python that auto-generates edge-case inputs. | github.com/HypothesisWorks/hypothesis |
| Nix | Purely functional package manager for reproducible, declarative environments. | github.com/NixOS/nix |
| GNU Guix | Reproducible, transactional package manager with a focus on scientific workflows. | guix.gnu.org |
| Pixi | Rust-based cross-platform package manager built on the conda ecosystem with workspace lockfiles. | github.com/prefix-dev/pixi |
Before running an experiment, an autonomous agent needs to know what's already been tried. This category covers literature APIs, citation graphs, and reference managers agents can query programmatically. Coverage, freshness, and rate limits vary — most real systems combine two or three sources.
| Tool | Description | Link |
|---|---|---|
| Semantic Scholar API | Free API from Allen AI over hundreds of millions of papers, with paper search, citations, and recommendations. | api.semanticscholar.org |
| arXiv API | Official query API over arXiv's preprint corpus, returning Atom/XML metadata. | info.arxiv.org/help/api |
| OpenAlex | Open catalog of the global research system with a free, no-auth REST API over works, authors, and venues. | openalex.org |
| Zotero | Open-source reference manager with a web API, browser connectors, and a Python client. | github.com/zotero/zotero |
| Paperpile | Cloud-based reference manager for Google Docs and the web with an API for library access. | paperpile.com |
| Connected Papers | Graph-based visual explorer that surfaces similar papers via co-citation and bibliographic coupling. | connectedpapers.com |
| Elicit | AI research assistant with paper search, data extraction, and systematic-review workflows. | elicit.com |
| Consensus | Evidence-focused search engine over peer-reviewed literature with a Consensus Meter for agreement. | consensus.app |
| ScholarAI | Research assistant over 200M+ papers and patents with citation generation and Zotero sync. | scholarai.io |
| Scite | Smart Citations platform classifying citations as supporting, contrasting, or mentioning. | scite.ai |
Most scientific knowledge is trapped in PDFs with tables, equations, and figures. These parsers extract structured content — some use layout models, some use LLMs, some are classic rule-based engines tuned for scientific papers. For high-throughput ingestion, benchmark on your actual corpus before committing.
| Library | Description | Link |
|---|---|---|
| Docling | IBM/LF AI open-source document parser with unified DocTags output for PDFs, slides, and images. | github.com/docling-project/docling |
| PyMuPDF | Python bindings to MuPDF for PDF parsing, rendering, and text and image extraction. | github.com/pymupdf/PyMuPDF |
| Crawl4AI | LLM-friendly web crawler that outputs clean markdown and structured data for ingestion. | github.com/unclecode/crawl4ai |
| GROBID | Java machine-learning library that structures scholarly PDFs into TEI XML with high accuracy on references. | github.com/kermitt2/grobid |
| Marker | Fast PDF and office-doc to Markdown/JSON converter with optional LLM-assisted refinement. | github.com/datalab-to/marker |
| Nougat | Meta's vision transformer for converting scientific PDFs to Markdown with equation support. | github.com/facebookresearch/nougat |
| tabula-py | Maintained Python wrapper around Tabula for extracting tables from text-based PDFs. | github.com/chezou/tabula-py |
| Camelot | Python library focused on extracting tables from PDFs with stream and lattice parsers. | github.com/camelot-dev/camelot |
| Unstructured | ETL library for partitioning, cleaning, and chunking 25+ document types for LLM pipelines. | github.com/Unstructured-IO/unstructured |
| pdfplumber | Pure-Python PDF parser with fine-grained access to characters, tables, and layout metadata. | github.com/jsvine/pdfplumber |
These are models and agent systems specifically designed for scientific reasoning, hypothesis generation, or domain-specific tasks (chemistry, biology, medicine). Some are open weights, some are research prototypes with released code, some are closed APIs. Expect the frontier to shift every few months.
| Model / System | Description | Link |
|---|---|---|
| Sakana AI Scientist v2 | End-to-end agentic system that generates ideas, runs experiments, and drafts papers via agentic tree search. | github.com/SakanaAI/AI-Scientist-v2 |
| FunSearch | DeepMind method pairing LLM program search with an evaluator, used to find new cap-set and bin-packing solutions. | github.com/google-deepmind/funsearch |
| ChemCrow | LangChain-based chemistry agent with 18 tool integrations for synthesis planning and molecule property lookup. | github.com/ur-whitelab/chemcrow-public |
| Coscientist | GPT-4-driven autonomous chemistry agent from the Gomes group, demonstrated on palladium cross-couplings. | github.com/gomesgroup/coscientist |
| DARWIN | Open 7B foundation model fine-tuned on physics, chemistry, and materials-science literature. | github.com/MasterAI-EAM/Darwin |
| Galactica | Meta's 120B scientific language model (2022); publicly retracted shortly after release due to hallucination issues — kept here as historical reference with weights on Hugging Face. | github.com/paperswithcode/galai |
| SciBERT | BERT variant trained on 1.14M scientific papers from Semantic Scholar, still used as a scientific-NLP baseline. | github.com/allenai/scibert |
| BioGPT | Microsoft biomedical GPT pretrained on PubMed abstracts, available via Hugging Face Transformers. | github.com/microsoft/BioGPT |
| Med-PaLM | Google Research family of medical LLMs evaluated on USMLE-style and clinical-reasoning benchmarks. | research.google med-palm |
| ESM3 | EvolutionaryScale's multimodal protein model for joint sequence, structure, and function generation. | github.com/evolutionaryscale/esm |
LLM-driven science agents reach wet labs faster when they pre-screen candidates in silico. The libraries here solve physics-based problems — molecular dynamics, electronic structure, finite-element PDEs, reaction kinetics — cheaply enough to rank thousands of hypotheses before committing reagent or instrument time in an autonomous loop.
| Library | Description | Link |
|---|---|---|
| LAMMPS | Classical molecular dynamics engine scaling from single workstations to GPU supercomputers. | github.com/lammps/lammps |
| GROMACS | Biomolecular MD package optimized for proteins, lipids, and drug-binding free-energy work. | github.com/gromacs/gromacs |
| PySCF | Python quantum chemistry package covering Hartree-Fock, DFT, MP2, and coupled cluster. | github.com/pyscf/pyscf |
| Psi4 | Open-source quantum chemistry program with a Python API for electronic-structure methods. | github.com/psi4/psi4 |
| ASE | Python toolkit wrapping 30+ DFT and MD calculators behind a uniform Atoms object. | gitlab.com/ase/ase |
| RDKit | Cheminformatics toolkit for molecular descriptors, substructure search, and SMARTS filters. | github.com/rdkit/rdkit |
| OpenMM | GPU-accelerated biomolecular MD engine with a first-class Python scripting API. | github.com/openmm/openmm |
| FEniCS / DOLFINx | Python FEM framework for solving PDEs via variational forms in UFL. | github.com/FEniCS/dolfinx |
| Cantera | Chemical kinetics, thermodynamics, and transport library for reactors and combustion. | github.com/Cantera/cantera |
| deal.II | C++ finite element library with Python bindings for adaptive-mesh PDE solvers. | github.com/dealii/dealii |
BO/AL is the decision core of closed-loop autonomous experimentation — given what has been observed so far, pick the next experiment to run under a limited budget. The libraries split along two lines: Gaussian-process-based frameworks (BoTorch, GPyTorch, Emukit, Trieste) that shine on smooth, low-dimensional design spaces, and algorithm-agnostic optimizers (Nevergrad, SMAC3, Hyperopt, Vizier) that scale to high-dimensional or conditional search spaces.
| Library | Description | Link |
|---|---|---|
| BoTorch | PyTorch library for Bayesian optimization with Monte Carlo acquisition functions. | github.com/pytorch/botorch |
| Ax | Meta platform built on BoTorch for adaptive experimentation and closed-loop tuning. | github.com/facebook/Ax |
| GPyTorch | Scalable Gaussian processes in PyTorch; the GP backbone underneath BoTorch. | github.com/cornellius-gp/gpytorch |
| Emukit | Decision-making toolbox covering BO, experimental design, and sensitivity analysis. | github.com/EmuKit/emukit |
| Hyperopt | Distributed hyperparameter optimization using Tree-structured Parzen Estimators. | github.com/hyperopt/hyperopt |
| SMAC3 | Sequential model-based algorithm configuration for AutoML and black-box tuning. | github.com/automl/SMAC3 |
| Nevergrad | Gradient-free optimization platform with evolutionary and population-based algorithms. | github.com/facebookresearch/nevergrad |
| Trieste | TensorFlow/GPflow-based BO library with trust-region and multi-fidelity strategies. | github.com/secondmind-labs/trieste |
| Vizier | Open-source release of Google's internal black-box optimization service. | github.com/google/vizier |
| scikit-activeml | Active learning library with pool-based query strategies on scikit-learn models. | github.com/scikit-activeml/scikit-activeml |
Reinforcement learning fits scientific problems where the agent must choose sequential experiments under uncertainty — active learning, Bayesian optimization loops, molecule design, reaction planning. The libraries below are general-purpose RL frameworks used as substrates in science RL work.
| Library | Description | Link |
|---|---|---|
| Gymnasium | Maintained fork of OpenAI Gym with the standard RL environment API and a large environment registry. | github.com/Farama-Foundation/Gymnasium |
| Stable-Baselines3 | PyTorch implementations of standard RL algorithms with a consistent, production-friendly API. | github.com/DLR-RM/stable-baselines3 |
| CleanRL | Single-file PyTorch RL implementations used widely for reproducible benchmarks and teaching. | github.com/vwxyzjn/cleanrl |
| TRL | Hugging Face library for RLHF, DPO, PPO, and GRPO fine-tuning of language models. | github.com/huggingface/trl |
| Ray RLlib | Distributed RL library inside Ray with a large algorithm zoo and offline-RL support. | github.com/ray-project/ray |
| Tianshou | PyTorch RL library with modular agents, on-policy and off-policy algorithms, and offline RL. | github.com/thu-ml/tianshou |
| Sample Factory | High-throughput async PPO library optimized for single-machine, multi-GPU training. | github.com/alex-petrenko/sample-factory |
| PettingZoo | Multi-agent environment API and environment zoo from the Farama Foundation. | github.com/Farama-Foundation/PettingZoo |
| PFRL | PyTorch-based deep RL library from Preferred Networks with a suite of modern algorithms. | github.com/pfnet/pfrl |
Retrieval-augmented agents need a vector store for papers, protocols, lab notes, and prior results. Pick based on scale and operational model — some run embedded, some are managed services, some are Postgres extensions. For embeddings, science-tuned models often beat general-purpose ones on domain corpora.
| Tool | Description | Link |
|---|---|---|
| Qdrant | Rust-based vector database with filtering, quantization, and a managed cloud option. | github.com/qdrant/qdrant |
| Weaviate | Vector database with built-in hybrid search, modules for embeddings, and GraphQL API. | github.com/weaviate/weaviate |
| Milvus | High-scale open-source vector database with GPU-accelerated index types. | github.com/milvus-io/milvus |
| Chroma | Embedded and server vector database focused on developer ergonomics for RAG apps. | github.com/chroma-core/chroma |
| Pinecone | Managed serverless vector database with hybrid search and metadata filtering. | pinecone.io |
| FAISS | Meta's library for efficient similarity search and clustering of dense vectors. | github.com/facebookresearch/faiss |
| LanceDB | Serverless vector database built on the Lance columnar format for multimodal search. | github.com/lancedb/lancedb |
| pgvector | PostgreSQL extension that adds vector types, HNSW/IVF indexing, and ANN search. | github.com/pgvector/pgvector |
| sentence-transformers | Python library for dense sentence and passage embeddings on top of Hugging Face models. | github.com/UKPLab/sentence-transformers |
| BAAI bge-m3 | Multi-lingual multi-function embedding model that handles dense, sparse, and ColBERT-style retrieval. | huggingface.co/BAAI/bge-m3 |
Lab instruments, compute jobs, and agents emit events that need routing, buffering, and durable replay. These brokers and streaming systems form the nervous system of a distributed autonomous lab. Choose on durability guarantees and operational fit, not raw throughput.
| Tool | Description | Link |
|---|---|---|
| NATS | High-performance cloud-native messaging system with JetStream for durable streams and KV. | github.com/nats-io/nats-server |
| Apache Kafka | Distributed log-based event streaming platform with an enormous connector ecosystem. | github.com/apache/kafka |
| RabbitMQ | AMQP-centric broker for traditional messaging patterns like work queues and routing. | github.com/rabbitmq/rabbitmq-server |
| Redis Streams | Append-only log data structure built into Redis with consumer groups. | redis.io streams |
| Apache Pulsar | Pub/sub and queue system with tiered storage and geo-replication built in. | github.com/apache/pulsar |
| ZeroMQ | Embeddable messaging library for low-latency socket-level patterns. | github.com/zeromq/libzmq |
| Celery | Python distributed task queue with a broker-based backend, common in lab automation backends. | github.com/celery/celery |
| Apache Flink | Stateful stream-processing engine with exactly-once semantics and event-time windows. | github.com/apache/flink |
| Apache Beam | Unified programming model for batch and streaming pipelines across multiple runners. | github.com/apache/beam |
| Redpanda | Kafka-API-compatible streaming platform written in C++ with no ZooKeeper dependency. | github.com/redpanda-data/redpanda |
When agents can call tools that move real money, compounds, or lab hardware, you need input filters, output validators, and policy engines. This category ranges from prompt-injection detectors to structured-output validators to full red-teaming toolkits. Layer several — no single tool covers all failure modes.
| Tool | Description | Link |
|---|---|---|
| Guardrails AI | Python framework for structured output validation and input/output guards via reusable validators. | github.com/guardrails-ai/guardrails |
| NeMo Guardrails | NVIDIA toolkit for adding programmable rails to LLM apps via the Colang dialog language. | github.com/NVIDIA/NeMo-Guardrails |
| LLM Guard | Protect AI's prompt and output scanner covering toxicity, PII, secrets, and prompt injection. | github.com/protectai/llm-guard |
| NVIDIA Garak | LLM vulnerability scanner that probes for jailbreaks, hallucination, and prompt injection. | github.com/NVIDIA/garak |
| DeepTeam | LLM red-teaming framework with 40+ vulnerability scanners and OWASP-aligned test suites. | github.com/confident-ai/deepteam |
| Lakera Guard | Commercial real-time API for detecting prompt injection, jailbreaks, and data exfiltration. | lakera.ai |
| Invariant | Rule-based contextual guardrails for LLM and MCP tool-calling deployed as a proxy. | github.com/invariantlabs-ai/invariant |
| LlamaFirewall | Meta's agent guardrail framework combining PromptGuard 2, alignment checks, and CodeShield. | github.com/meta-llama/PurpleLlama |
| PyRIT | Microsoft's Python risk-identification toolkit for automated adversarial testing of generative AI. | github.com/microsoft/PyRIT |
Beyond traditional infra monitoring, autonomous-science systems need traces across agent steps, tool calls, and LLM responses. The first group here covers generic infra observability; the second is LLM-specific. You want both — one tells you the cluster is healthy, the other tells you the agent is sane.
| Tool | Description | Link |
|---|---|---|
| Prometheus | Pull-based time-series monitoring system with a powerful query language and alerting. | github.com/prometheus/prometheus |
| Grafana | Open-source dashboarding tool that plots metrics, logs, and traces across datasources. | github.com/grafana/grafana |
| Loki | Horizontally scalable log aggregation system with a label-based query model similar to Prometheus. | github.com/grafana/loki |
| OpenTelemetry | CNCF standard for traces, metrics, and logs with SDKs and collectors across languages. | github.com/open-telemetry |
| Jaeger | CNCF distributed tracing platform for investigating latency in microservice architectures. | github.com/jaegertracing/jaeger |
| Helicone | Open-source LLM observability and AI gateway with tracing, costs, and prompt management. | github.com/Helicone/helicone |
| LangSmith | LangChain's hosted platform for agent tracing, evaluation, and prompt management. | langchain.com/langsmith |
| Arize Phoenix | OpenTelemetry-based AI observability with tracing, eval, datasets, and experiments. | github.com/Arize-ai/phoenix |
| Opik | Comet's open-source LLM observability tool with tracing, automated evaluations, and dashboards. | github.com/comet-ml/opik |
| Langfuse | Self-hostable LLM engineering platform with tracing, prompt management, and datasets. | github.com/langfuse/langfuse |
Contributions are welcome. See CONTRIBUTING.md for the rules on what belongs, what doesn't, and how to submit a PR.
Released under the MIT License.
Built and maintained by Scivity Labs — building the operating system for autonomous science.
If you find this useful, please ⭐ star the repository.