Platforms / Packages / Protocols / Systems / Operations

The AI Enabling Stack

A comprehensive, dated landscape of 428 AI-enabling tools, platforms, protocols, packages, and operational systems. The selected field guide explains how 89 important tools behave; the whole-landscape index maps 339 additional options you may need to recognize, compare, or call.

428 total options89 detailed field guides18 landscape categoriesVerified Aug 28, 2026

The classification rule

Models think.
Stacks make them act.

AI-native software is designed around models or agents. AI-focused software specializes an existing engineering function for AI workloads. General infrastructure becomes part of the AI system when a model can call it reliably through code, a CLI, an API, a protocol, or a user interface.

Selected field guide / operational detail

How the important tools actually get used.

89 fully described tools, including 11 recovered from Cory's bookmarks. These entries specify the call surface, unlocked capability, and the conditions where each tool earns its place.

Search capabilitiesCapability layerRelationshipProvenance

89 detailed tools shown

01 / Agent frameworks & orchestration

12 tools

AI-nativePython SDK

OpenAI Agents SDK

Runs tool-using agents with handoffs, guardrails, sessions, tracing, and structured outputs.

How AI calls it: Import Agent and Runner; register Python functions or MCP tools.
Use it when: You want a compact agent loop with explicit tools and production controls.

Primary-source research

AI-nativePython / TypeScript SDK

LangGraph

Models long-running agent behavior as stateful graphs with durable execution and interrupts.

How AI calls it: Define graph nodes and edges, compile, then invoke or stream the graph.
Use it when: The workflow needs branches, checkpoints, resumability, or human approval.

Primary-source research

AI-nativePython SDK / CLI

CrewAI

Coordinates specialized agents, crews, tasks, and event-driven flows.

How AI calls it: Define Agent, Task, Crew, or Flow objects and run them through the CLI or Python.
Use it when: Roles and delegated multi-agent collaboration are central to the design.

Primary-source research

AI-nativePython / .NET SDK

Microsoft AutoGen

Builds message-driven single- and multi-agent systems with distributed runtimes and code execution.

How AI calls it: Install AutoGen packages, define routed agents, and run them in a local or distributed runtime.
Use it when: You need explicit inter-agent messaging or distributed agent execution.

Primary-source research

AI-nativePython / .NET / Java SDK

Semantic Kernel

Connects models to plugins, planners, memory, and enterprise application code.

How AI calls it: Register plugins and services in a Kernel, then invoke functions or agents.
Use it when: AI orchestration must fit strongly typed Microsoft or enterprise applications.

Primary-source research

AI-nativePython SDK

Pydantic AI

Creates type-safe agents with dependency injection, tools, structured results, and eval support.

How AI calls it: Declare an Agent with a model, result type, dependencies, and decorated tools.
Use it when: Schema correctness and Python typing should govern agent behavior.

Primary-source research

AI-nativeTypeScript SDK / CLI

Mastra

Combines agents, typed tools, workflows, memory, MCP, tracing, and evaluations.

How AI calls it: Create TypeScript agents and tools, then compose them into resumable workflows.
Use it when: You want a full agent stack in the TypeScript ecosystem.

Primary-source research

AI-nativePython package

smolagents

Provides lightweight tool-calling and code-writing agents with Hugging Face integration.

How AI calls it: Instantiate ToolCallingAgent or CodeAgent with models and registered tools.
Use it when: You want a small, inspectable agent framework or code-agent prototype.

Primary-source research

AI-nativePython / TypeScript SDK

LlamaIndex Workflows

Builds event-driven AI workflows around retrieval, tools, and data agents.

How AI calls it: Define event handlers as workflow steps and run or stream the workflow.
Use it when: The agent is primarily organized around documents, indexes, and retrieval.

Primary-source research

AI-focusedVisual workflow / API

n8n

Connects AI agents and models to application triggers, credentials, APIs, and deterministic workflow nodes.

How AI calls it: Assemble nodes visually, expose webhooks, or call workflows through the API.
Use it when: Business automation needs AI inside a visible, inspectable workflow.

Bookmark + primary-source research

AI-nativeVisual builder / API

Flowise

Builds agent, retrieval, and tool-calling flows through visual graphs backed by deployable APIs.

How AI calls it: Compose nodes in the visual builder, then call the published prediction API.
Use it when: Teams need to prototype and expose AI flows without hand-coding every integration.

Bookmark + primary-source research

AI-nativeVisual platform / API

Dify

Combines prompt applications, agents, workflows, retrieval, tools, evaluations, and hosted APIs.

How AI calls it: Build an application visually and invoke it through Dify's generated API.
Use it when: A cross-functional team needs one surface for building and operating AI applications.

Primary-source research

02 / Inference & local runtimes

13 tools

AI-nativePython package / server

vLLM

Serves large language models with high-throughput batching, parallelism, quantization, and OpenAI-compatible APIs.

How AI calls it: Run the vLLM server CLI or import its Python inference engine.
Use it when: You need production-scale open-model inference on GPUs.

Primary-source research

AI-nativeC/C++ library / CLI / server

llama.cpp

Runs quantized language and multimodal models efficiently across CPUs and many accelerator backends.

How AI calls it: Use llama-cli, llama-server, or link the native library.
Use it when: Portable local inference, GGUF models, or minimal runtime dependencies matter.

Primary-source research

AI-nativeDesktop daemon / CLI / API

Ollama

Downloads, runs, and exposes local models through a simple command line and HTTP API.

How AI calls it: Run ollama pull/run or call the local generate and chat endpoints.
Use it when: Developers need the fastest route to repeatable local-model use.

Primary-source research

AI-focusedPython / C++ framework

MLX

Runs and trains array-based ML workloads efficiently on Apple silicon unified memory.

How AI calls it: Install mlx, load arrays or models, and execute Python functions on Apple GPU/CPU.
Use it when: Local AI work targets Macs and Apple silicon.

Primary-source research

AI-nativePython / C++ runtime

TensorRT-LLM

Compiles and serves LLMs with NVIDIA-specific kernels, quantization, batching, and parallelism.

How AI calls it: Build an optimized engine and run it through the runtime or serving stack.
Use it when: NVIDIA GPU latency and throughput require aggressive optimization.

Primary-source research

AI-nativeCross-platform runtime / APIs

ONNX Runtime GenAI

Runs generative models in the ONNX ecosystem across cloud, desktop, mobile, and edge hardware.

How AI calls it: Load an ONNX model and configure generation through native or language bindings.
Use it when: One model runtime must span different hardware and deployment targets.

Primary-source research

AI-nativeContainer / HTTP server

Text Generation Inference

Hosts Hugging Face text-generation models with streaming, tensor parallelism, and production telemetry.

How AI calls it: Launch its Docker image or launcher and call the HTTP API.
Use it when: Hugging Face models need a standardized production server.

Primary-source research

AI-nativeDesktop app / CLI / API

LM Studio

Discovers and runs local models with chat, developer tools, and OpenAI-compatible endpoints.

How AI calls it: Load a model in the app or CLI and start the local API server.
Use it when: A visual local-model workstation is preferable to assembling runtimes manually.

Primary-source research

AI-nativePython SDK / proxy

LiteLLM

Normalizes model-provider APIs, routing, retries, budgets, logging, and fallbacks behind an OpenAI-compatible interface.

How AI calls it: Call providers through the Python client or deploy the LiteLLM proxy.
Use it when: Applications must switch or route across many model providers.

Primary-source research

AI-focusedPython SDK / service runtime

BentoML

Packages model inference and supporting code into scalable API services and containers.

How AI calls it: Declare Bento services and APIs, then serve locally or build deployable images.
Use it when: Custom inference code needs a production service boundary.

Primary-source research

AI-focusedSelf-hosted web application / API

Open WebUI

Provides a local-first interface for models, tools, knowledge, functions, and multi-provider chat.

How AI calls it: Deploy the container, connect model endpoints, and expose tools or pipelines through its interfaces.
Use it when: Teams need a controlled self-hosted workspace for local and remote models.

Primary-source research

AI-focusedDesktop / self-hosted application / API

AnythingLLM

Combines local or hosted models with document workspaces, retrieval, agents, and tools.

How AI calls it: Install the desktop or server edition, connect providers, and create document-backed workspaces.
Use it when: A ready-made private knowledge and agent workspace is more useful than a framework.

Primary-source research

AI-nativeResearch repository

TinyRecursiveModels

Provides a compact recursive-reasoning model implementation and reference training code.

How AI calls it: Clone the repository and run its published training or evaluation scripts.
Use it when: You want to study small-model recursive reasoning rather than only consume hosted APIs.

Cory's bookmarks

03 / Retrieval, memory & databases

10 tools

AI-nativeVector database / clients

Qdrant

Stores vectors and payloads for filtered semantic, hybrid, sparse, and multi-vector retrieval.

How AI calls it: Create collections and query them through REST, gRPC, or official SDKs.
Use it when: RAG or agent memory needs a purpose-built vector service.

Primary-source research

AI-nativeManaged vector database / SDK

Pinecone

Provides hosted dense, sparse, and hybrid retrieval with metadata filtering.

How AI calls it: Create an index, upsert embeddings, and query it through an SDK or API.
Use it when: Managed vector search is preferred over operating infrastructure.

Primary-source research

AI-nativeVector database / GraphQL / clients

Weaviate

Combines object storage, vector search, hybrid retrieval, filters, and optional model integrations.

How AI calls it: Define collections and call its client, REST, or GraphQL APIs.
Use it when: Semantic retrieval needs flexible schemas and integrated vectorization.

Primary-source research

AI-nativeDistributed vector database

Milvus

Indexes and searches large vector collections with multiple index strategies and distributed scaling.

How AI calls it: Create collections and search them using pymilvus or another official SDK.
Use it when: Vector search volume or scale exceeds a simple embedded store.

Primary-source research

General infrastructurePostgreSQL extension

pgvector

Adds vector columns, distance operators, and approximate indexes to PostgreSQL.

How AI calls it: Enable the extension and issue SQL vector similarity queries.
Use it when: Embeddings should remain beside relational source data.

Primary-source research

AI-nativeEmbedded/server vector database

Chroma

Stores documents, metadata, and embeddings behind a developer-friendly retrieval API.

How AI calls it: Create collections through Python or JavaScript and add/query records.
Use it when: A prototype or local RAG application needs minimal setup.

Primary-source research

AI-focusedEmbedded/cloud multimodal database

LanceDB

Stores vectors and multimodal data on the Lance columnar format with hybrid search.

How AI calls it: Connect with Python, TypeScript, or Rust; create tables and run vector queries.
Use it when: Local-first or object-storage-backed retrieval should stay lightweight.

Primary-source research

General infrastructureIn-memory database / query engine

Redis Vector Search

Combines vector similarity, metadata filters, caching, streams, and fast operational state.

How AI calls it: Create vector indexes and issue FT.SEARCH queries through Redis clients.
Use it when: Agent state, cache, queues, and retrieval should share one low-latency system.

Primary-source research

AI-focusedC++ / Python library

FAISS

Performs efficient similarity search and clustering over dense vectors locally.

How AI calls it: Build an index in Python or C++, add vectors, and search nearest neighbors.
Use it when: Retrieval runs in-process and does not need database semantics.

Primary-source research

General infrastructureGraph database / Cypher

Neo4j Vector Index

Combines vector similarity with graph relationships and structured traversal.

How AI calls it: Create vector indexes in Cypher and query nodes or relationships by similarity.
Use it when: Retrieval quality depends on connected entities and knowledge graphs.

Primary-source research

04 / Documents, OCR & ingestion

10 tools

AI-focusedPython package / CLI / API / MCP

Docling

Converts PDFs, Office files, HTML, images, and audio into structured document representations.

How AI calls it: Run the docling CLI, import DocumentConverter, or call its service/MCP interface.
Use it when: RAG ingestion must preserve layout, tables, reading order, and provenance.

Primary-source research

AI-focusedPython library / API

Unstructured

Partitions heterogeneous enterprise documents into typed elements for downstream chunking and retrieval.

How AI calls it: Call partition functions locally or send files to the hosted API.
Use it when: A broad file-ingestion layer must normalize many document formats.

Primary-source research

AI-nativeManaged parsing API / SDK

LlamaParse

Turns complex documents into structured text and markdown optimized for LLM retrieval.

How AI calls it: Submit documents through the LlamaCloud SDK or REST API.
Use it when: Scanned, table-heavy, or irregular documents defeat conventional parsers.

Primary-source research

General infrastructurePython package

PyMuPDF

Extracts, renders, searches, annotates, and transforms PDF and document content.

How AI calls it: Open documents with fitz/PyMuPDF and call page-level text, image, or geometry APIs.
Use it when: AI pipelines need fast programmatic access to native PDF structure.

Primary-source research

General infrastructureJavaScript package

PDF.js / pdfjs-dist

Parses and renders PDFs in browsers or Node-based document workflows.

How AI calls it: Install pdfjs-dist and load documents through the PDF.js API.
Use it when: An agent-facing web app must inspect, render, or map PDF pages.

Bookmark + primary-source research

General infrastructureNative CLI / library

Tesseract OCR

Recognizes text in images and scanned pages across many languages.

How AI calls it: Run the tesseract CLI or call its native and wrapper APIs.
Use it when: A local, inspectable OCR baseline is sufficient.

Primary-source research

AI-focusedPython toolkit / CLI

PaddleOCR

Performs multilingual OCR, layout analysis, table recognition, and key-information extraction.

How AI calls it: Install PaddleOCR and call its pipelines from Python or the command line.
Use it when: Documents need modern multilingual recognition and structural parsing.

Primary-source research

General infrastructureCLI / Python package

OCRmyPDF

Adds searchable OCR text layers to scanned PDFs while preserving the original pages.

How AI calls it: Run ocrmypdf against input and output files or call its Python API.
Use it when: Scanned archives should become searchable before AI ingestion.

Primary-source research

AI-focusedPython package / CLI

MarkItDown

Converts common files and media into Markdown for indexing and model context.

How AI calls it: Run markitdown from the CLI or call its Python converter.
Use it when: Simple, readable Markdown is the preferred interchange format.

Primary-source research

General infrastructurePython / native service

Gentle forced aligner

Aligns transcript words to exact audio timestamps for searchable or editable media.

How AI calls it: Run its local service or command-line alignment pipeline against audio and text.
Use it when: An AI workflow needs time-coded transcripts, captions, or clip selection.

Cory's bookmarks

05 / Browser, computer & web automation

10 tools

General infrastructureNode / Python / Java / .NET SDK

Playwright

Automates Chromium, Firefox, and WebKit with DOM, network, file, screenshot, and input control.

How AI calls it: Launch a browser and call page actions through its SDK or CLI.
Use it when: An agent must operate or verify a real web interface reliably.

Primary-source research

General infrastructureWebDriver SDK

Selenium

Controls browsers through a standardized cross-language automation protocol.

How AI calls it: Create a WebDriver session and issue navigation, locator, and input commands.
Use it when: Browser automation must fit established WebDriver infrastructure.

Primary-source research

AI-nativePython package / cloud API

Browser Use

Wraps browser state and actions into an agent-oriented interface for language models.

How AI calls it: Create an Agent with a browser session, model, task, and allowed tools.
Use it when: A model should decide browser actions rather than follow a fixed script.

Primary-source research

AI-focusedAPI / SDK / MCP

Firecrawl

Crawls, searches, and extracts websites into clean markdown or structured data for AI systems.

How AI calls it: Call scrape, crawl, map, search, or extract endpoints through an SDK or MCP.
Use it when: Agents need web content without maintaining a browser/crawler stack.

Primary-source research

AI-focusedPython package / server

Crawl4AI

Runs asynchronous browser crawling and produces LLM-oriented markdown and structured extraction.

How AI calls it: Configure AsyncWebCrawler or deploy its local server and call it programmatically.
Use it when: Web ingestion should stay local, open source, and model-aware.

Primary-source research

General infrastructurePython crawling framework

Scrapy

Builds high-throughput spiders, extraction pipelines, throttling, retries, and crawl scheduling.

How AI calls it: Define Spider classes and run them through the scrapy CLI or API.
Use it when: Large deterministic crawls matter more than interactive browser behavior.

Primary-source research

AI-nativeTypeScript / Python SDK

Stagehand

Combines deterministic browser code with model-directed act, extract, and observe primitives.

How AI calls it: Create a Stagehand session and call act, extract, or observe from code.
Use it when: Automation should blend reliable selectors with semantic AI control.

Primary-source research

AI-focusedHosted actors / API / SDK

Apify

Runs packaged web automations, crawlers, datasets, schedules, proxies, and integrations.

How AI calls it: Start Actors through the console, API, CLI, SDK, or MCP integration.
Use it when: Reusable browser/crawler workloads need managed execution and storage.

Primary-source research

General infrastructureNode.js SDK

Puppeteer

Controls Chrome and Firefox through a high-level JavaScript API over browser protocols.

How AI calls it: Launch a browser, open pages, and drive DOM or network actions from Node.
Use it when: Chrome-centric automation belongs directly inside a JavaScript service.

Primary-source research

General infrastructureBrowser protocol

Chrome DevTools Protocol

Exposes low-level browser domains for DOM, network, runtime, storage, tracing, and debugging control.

How AI calls it: Connect to a Chrome debugging target and send protocol commands over WebSocket.
Use it when: An agent tool needs control below the abstractions of browser libraries.

Primary-source research

06 / Media & multimodal pipelines

14 tools

General infrastructureCLI / native libraries

FFmpeg

Decodes, transforms, filters, streams, and encodes audio, video, images, subtitles, and metadata.

How AI calls it: Call ffmpeg/ffprobe commands or link libav* libraries.
Use it when: AI-generated or analyzed media must be normalized, clipped, composited, or delivered.

Primary-source research

General infrastructureC++ / Python / Java library

OpenCV

Processes images and video for geometry, tracking, detection support, camera input, and visualization.

How AI calls it: Import cv2 or link OpenCV and call frame/image operations.
Use it when: A vision pipeline needs deterministic operations around a model.

Primary-source research

AI-nativePython package / CLI

Whisper

Transcribes and translates multilingual speech with open speech-recognition models.

How AI calls it: Load a Whisper model in Python or run the whisper CLI against media.
Use it when: Audio or video must become searchable text and timestamps.

Primary-source research

AI-focusedPython package

faster-whisper

Runs Whisper models through CTranslate2 with lower memory use and faster inference.

How AI calls it: Instantiate WhisperModel and transcribe audio from Python.
Use it when: Whisper throughput or deployment efficiency is a bottleneck.

Primary-source research

AI-nativePython package

Diffusers

Provides reusable diffusion pipelines, schedulers, models, adapters, and training utilities.

How AI calls it: Load a pretrained pipeline and call it with prompts, images, or control inputs.
Use it when: Code must generate or transform images, video, or audio with diffusion models.

Primary-source research

AI-nativeNode graph / server API

ComfyUI

Builds inspectable generative-media graphs from models, samplers, conditioning, control, and post-processing nodes.

How AI calls it: Connect nodes visually, submit workflow JSON, or call its server API.
Use it when: Complex generation pipelines need visible graphs and reusable workflows.

Primary-source research

AI-nativeModel architecture / implementations

ControlNet

Adds spatial controls such as edges, depth, pose, and segmentation to diffusion generation.

How AI calls it: Load a compatible ControlNet model through Diffusers or a node-based UI.
Use it when: Generated imagery must follow an explicit composition or structure.

Primary-source research

AI-focusedPython package / CLI / server

rembg

Removes image backgrounds using segmentation models in batch, library, or service modes.

How AI calls it: Run rembg from the CLI, import remove(), or start its HTTP server.
Use it when: An agent needs transparent cutouts or subject isolation as a deterministic step.

Primary-source research

AI-nativeWeb application

Topaz Astra

Upscales and enhances generative or low-resolution video for creative delivery.

How AI calls it: Upload video through the application and select an enhancement workflow.
Use it when: Generated video needs resolution and detail recovery before publishing.

Cory's bookmarks

AI-nativeWeb application

Topaz Bloom

Generates and refines high-resolution visual assets through a hosted creative workspace.

How AI calls it: Open the public Bloom workspace and submit image-generation or refinement tasks.
Use it when: Creative work needs iterative high-resolution image generation in a visual tool.

Cory's bookmarks

AI-nativeAPI / SDK / web application

ElevenLabs

Generates speech, voices, sound effects, music, dubbing, and conversational audio.

How AI calls it: Call speech/audio APIs through official SDKs or use the web tools.
Use it when: An agent must create, transform, or converse through audio.

Bookmark + primary-source research

AI-nativeWeb application

OpenArt

Creates and edits images through prompt, reference, style, and workflow-driven generative tools.

How AI calls it: Use its public creative interface and model/workflow controls.
Use it when: A visual operator needs a broad hosted image-generation workspace.

Cory's bookmarks

AI-nativeWeb application

Hailuo AI Video

Generates video from text and image prompts using hosted video models.

How AI calls it: Submit prompts and source images through its generation interface.
Use it when: A creative workflow needs rapid hosted image-to-video or text-to-video generation.

Cory's bookmarks

AI-nativeWeb application

Higgsfield

Creates cinematic images and video with camera, style, character, and storyboard controls.

How AI calls it: Use the hosted creative workspace and its prompt-assistance controls.
Use it when: Generative video needs deliberate cinematic direction rather than a single prompt.

Cory's bookmarks

07 / Evaluation, tracing & safety

10 tools

AI-nativeSDK / observability platform

Langfuse

Captures LLM traces, prompts, scores, datasets, experiments, and production metrics.

How AI calls it: Instrument model and agent calls with its SDK or OpenTelemetry integrations.
Use it when: Teams need self-hostable evidence of what an AI system did and why.

Primary-source research

AI-nativeSDK / hosted platform

LangSmith

Traces agent runs, manages datasets, evaluates outputs, and monitors deployments.

How AI calls it: Wrap or instrument calls, then run evaluation experiments against datasets.
Use it when: LangChain/LangGraph systems need integrated debugging and evaluation.

Primary-source research

AI-nativeCLI / Node package

promptfoo

Runs repeatable prompt, model, agent, red-team, and security evaluations from declarative test cases.

How AI calls it: Define providers, prompts, assertions, and cases in config; run promptfoo eval or redteam.
Use it when: AI behavior must pass CI-style acceptance and adversarial tests.

Primary-source research

AI-nativePython package

RAGAS

Evaluates RAG and agent systems using metrics, test-set generation, and experiment workflows.

How AI calls it: Create evaluation datasets and call evaluate() with selected metrics.
Use it when: Retrieval quality and grounded answers need measurable regression tests.

Primary-source research

AI-nativePython testing framework

DeepEval

Expresses LLM and agent quality checks as unit tests with model-based and deterministic metrics.

How AI calls it: Write pytest-style test cases and run deepeval test or evaluate().
Use it when: AI evaluation should live beside application tests and CI.

Primary-source research

AI-nativeOpen-source observability / SDK

Arize Phoenix

Traces LLM applications and analyzes retrieval, embeddings, evaluations, and production behavior.

How AI calls it: Instrument with OpenTelemetry or SDK callbacks and inspect runs in Phoenix.
Use it when: Teams want open-source tracing plus embedding and retrieval diagnostics.

Primary-source research

AI-focusedPython platform

MLflow Evaluate

Logs model and GenAI evaluation results, metrics, artifacts, judges, and experiment comparisons.

How AI calls it: Call MLflow evaluation APIs inside a tracked experiment.
Use it when: AI evaluation belongs inside a broader model and experiment lifecycle.

Primary-source research

General infrastructureTelemetry standard / SDKs

OpenTelemetry

Carries vendor-neutral traces, metrics, logs, context, and semantic conventions across AI services.

How AI calls it: Instrument applications with language SDKs and export OTLP to a backend.
Use it when: AI traces must correlate with the rest of the production system.

Primary-source research

AI-nativePython / testing platform

Giskard

Scans and tests ML and LLM systems for performance, hallucination, bias, security, and robustness issues.

How AI calls it: Wrap a model and dataset, run scans, and convert findings into regression tests.
Use it when: Risk discovery should produce repeatable tests rather than one-time review.

Primary-source research

AI-nativePython / server

Guardrails AI

Validates and constrains model inputs and outputs with reusable validators and structured schemas.

How AI calls it: Define a Guard or validator set and wrap model calls or expose a guard server.
Use it when: Outputs must satisfy explicit safety, format, or quality constraints before use.

Primary-source research

08 / Deployment, compute & ML operations

10 tools

General infrastructureContainer runtime / CLI

Docker

Packages models, dependencies, tools, and services into reproducible isolated images.

How AI calls it: Build an image from a Dockerfile and run or compose containers.
Use it when: AI workloads must behave consistently across developer and server environments.

Primary-source research

General infrastructureCluster orchestrator / API

Kubernetes

Schedules, scales, networks, updates, and heals containerized AI services and jobs.

How AI calls it: Apply workload manifests or call the Kubernetes API through clients/controllers.
Use it when: Multiple services or GPU workloads require resilient cluster operations.

Primary-source research

AI-focusedPython distributed runtime

Ray

Distributes Python tasks, training, data processing, tuning, and model serving across clusters.

How AI calls it: Decorate tasks/actors with ray.remote or deploy applications through Ray Serve.
Use it when: Python AI workloads must scale beyond one process or machine.

Primary-source research

AI-focusedPython serverless platform

Modal

Runs GPU inference, training, batch jobs, sandboxes, and notebooks with infrastructure defined in code.

How AI calls it: Decorate Python functions/classes and launch them with the Modal CLI.
Use it when: GPU or sandbox workloads need rapid elastic execution without cluster management.

Primary-source research

AI-focusedGPU cloud / serverless API

Runpod

Provides GPU pods and autoscaling serverless endpoints for training and inference workloads.

How AI calls it: Launch pods or deploy endpoint workers through its API, CLI, or console.
Use it when: Dedicated or serverless GPUs are needed without owning hardware.

Primary-source research

AI-focusedPython / tracking and model platform

MLflow

Tracks experiments, packages models, manages registries, evaluates GenAI, and supports deployment workflows.

How AI calls it: Log runs and artifacts through the SDK, then register or serve models.
Use it when: Model and prompt changes need lineage across development and production.

Primary-source research

AI-focusedCLI / data versioning

DVC

Versions datasets, pipelines, parameters, metrics, and remote artifacts alongside Git.

How AI calls it: Track data with dvc add and define reproducible stages in dvc.yaml.
Use it when: Training or evaluation data must be reproducible without committing large files to Git.

Primary-source research

AI-focusedSDK / hosted platform

Weights & Biases

Tracks experiments, artifacts, evaluations, prompts, traces, and model-development workflows.

How AI calls it: Initialize runs with the SDK and log metrics, artifacts, tables, or traces.
Use it when: Teams need collaborative visibility across training and GenAI experiments.

Primary-source research

General infrastructurePython workflow orchestrator

Prefect

Schedules, retries, observes, and deploys data or AI workflows expressed as Python flows and tasks.

How AI calls it: Decorate Python functions, deploy flows, and run workers or managed execution.
Use it when: AI pipelines need durable scheduling and retries without heavy DAG ceremony.

Primary-source research

General infrastructurePython workflow scheduler

Apache Airflow

Coordinates scheduled batch pipelines as observable DAGs with extensive operators and integrations.

How AI calls it: Define DAGs and tasks in Python and run them through Airflow schedulers/workers.
Use it when: Established data-platform operations should govern recurring AI pipelines.

Primary-source research

Whole landscape / option index

Know the rest of the map.

A comprehensive, versioned map of meaningful software and systems an AI application or agent can call, use, run inside, or be built upon. It is not a ranking and cannot be permanently exhaustive. It includes commercial and open-source choices, standards, foundational packages, interfaces, hardware, and adjacent systems—not only products marketed as AI.

Search all optionsLandscape category

339 additional options shown

01 / Model APIs & intelligence platforms

24 optionsHosted foundation models, managed inference, and enterprise AI platforms.

Model API

Models think.Stacks make them act.

How the important tools actually get used.

Know the rest of the map.

OpenAI API

Anthropic API

Google AI for Developers

Google Vertex AI

Azure AI Foundry

Amazon Bedrock

Cohere

xAI API

GroqCloud

Together AI

Fireworks AI

Cerebras Inference

OpenRouter

Replicate

Hugging Face Inference Providers

Baseten

DeepInfra

Cloudflare Workers AI

NVIDIA NIM

IBM watsonx.ai

Databricks Mosaic AI

Snowflake Cortex AI

SambaNova Cloud

Perplexity API

OpenAI Codex

Claude Code

GitHub Copilot

Gemini CLI

Cursor

Windsurf

Cline

Roo Code

Kilo Code

Aider

OpenCode

OpenHands

Devin

Replit Agent

Amazon Q Developer

JetBrains Junie

Sourcegraph Amp

Continue

Tabnine

Qodo

Augment Code

Zed AI

Google Agent Development Kit

Agno

Haystack

DSPy

Vercel AI SDK

BeeAI Framework

CAMEL-AI

Langroid

TaskWeaver

Letta

Rasa

Models think.
Stacks make them act.