Environment, Compliance & Runtimes

AI-powered development environment where agents can write, run, and deploy code within sandboxed Repls. Natively integrates agentic code execution with cloud deployment. Ideal for demos and rapid prototyping by non-technical users.

AWS's open-source microVM technology powering Lambda and Fargate. Creates lightweight, KVM-based VMs in under 125ms for strong isolation. Used to build custom secure sandboxes for AI agent code execution at scale.

CodeSandbox's SDK for programmatically creating and managing cloud development environments. Used by AI coding agents to run and test generated code in isolated containers. Each sandbox gets its own filesystem, processes, and network.

Docker's official MCP toolkit for running MCP servers as isolated Docker containers. Provides a catalog of containerized tools and a secure runtime for agent tool execution. Integrates with Docker Desktop for local development.

Full Linux desktop environment accessible via API, enabling agents to control a browser, open applications, and interact with a GUI. Streams the desktop screen as a video or screenshots for visual grounding. Built on top of E2B's cloud infrastructure.

High-performance, general-purpose code execution engine supporting 70+ languages. Designed for low-latency execution with process isolation. Particularly well-suited for agents that need to run one-off code snippets in response to user queries.

Cloud sandbox infrastructure for running AI-generated code safely. Provides fast-booting microVMs with Python, Node.js, and custom environment support. REST and WebSocket APIs for real-time interaction. Used by leading AI coding assistants including Cursor and Claude Computer Use.

Isolated Python execution environments tailored for data science agents. Provides a secure REPL with pre-installed scientific libraries, persistent state between turns, and rich output rendering — enabling agents to run data analysis, generate charts, and process files safely.

Self-hosted container runtimes for agent code execution with full isolation guarantees. Running agent-generated code inside Docker containers provides strong security boundaries, reproducible environments, and easy cleanup — the baseline infrastructure pattern for on-prem agent deployments.

Secure, elastic infrastructure platform for AI-generated code execution and development environments. Provisions reproducible workspaces on any cloud or self-hosted. MCP server, CLI, and SDK available. Strong focus on security isolation for multi-tenant AI code execution.

Boot ephemeral micro-VMs in milliseconds for transient agent workloads. Fly Machines provide low-latency compute anywhere in the world with automatic cleanup after use — ideal for agents that need brief bursts of isolated compute without managing a persistent server fleet.

Instant cloud development environments with microVM isolation. AI-optimized API for spawning sandboxes, running code, and capturing output. Supports 1000+ templates and can warm up VMs to eliminate cold starts. Used by many coding agents as their execution environment.

Dagger project enabling AI coding agents to work in isolated containerized environments. Each agent gets its own container with full filesystem, preventing agents from interfering with each other. Native MCP integration for popular agent frameworks.

Real-time API traffic visibility tool for Kubernetes environments. Lets you monitor and inspect all network traffic flowing between agent microservices — essential for debugging unexpected agent behaviors, auditing tool calls, and ensuring compliance in production Kubernetes deployments.

Python runtime compiled to WebAssembly for running Python in browsers and Node.js. Execute Python code entirely client-side without a backend. Supports NumPy, pandas, matplotlib, and many scientific libraries. Ideal for sandboxed code execution without server infrastructure.

Secure JavaScript and TypeScript runtime with explicit permission model. Run untrusted code with fine-grained controls over filesystem, network, and environment access. Deploy on Deno Deploy for globally distributed edge execution of agent-generated code snippets.

Serverless cloud platform for running Python code, including GPU workloads. Instant scale-out with zero DevOps. Excellent for agent workloads that need burst compute, model inference, or parallel task execution with pay-per-use pricing.

Open-source online code execution system supporting 60+ programming languages. REST API for submitting code, managing execution queues, and retrieving results. Designed for competitive programming judges but widely used for secure agent code execution environments.

AI agent integrated directly into Replit's cloud development environment. Reads existing code, plans changes, installs packages, edits files, and runs the application — all within Replit's secure container. No local setup required, works entirely in the browser.

Serverless compute platform with millisecond billing and 1ms granularity. Each function invocation runs in an isolated execution environment. Support for custom container images enables arbitrary language runtimes. Widely used for sandboxing agent tool execution with IAM permission boundaries.

Open-source JavaScript/WASM runtime powering Cloudflare Workers. V8-based with strict isolation between worker instances. Can be run locally for testing. Extremely fast cold starts (~0ms) make it ideal for high-throughput agent action execution at the edge.

Cloud hosting platform for deploying web services, background workers, and cron jobs with a simple configuration model. First-class support for Docker containers makes it easy to deploy agent services. Free tier suitable for prototyping.

Deployment platform for deploying agent backends with zero configuration. Supports any language and framework with automatic deploys from Git. Offers persistent volumes and cron job scheduling for background agent processes.

Global GPU cloud with Serverless endpoints for deploying custom AI models. Supports any Docker container with auto-scaling from zero to thousands of instances. Popular for hosting open-source LLMs as inference backends for AI agents.

Serverless ML platform for deploying custom AI models with cold start times under 3 seconds. Deploy any Python ML function as a scalable API endpoint. Supports continuous batching for high-throughput LLM inference.

GPU cloud provider focused on AI and deep learning workloads with H100 and A100 clusters. Offers on-demand and reserved instances with persistent storage. Known for providing some of the most cost-effective H100 access for fine-tuning runs.

Peer-to-peer GPU marketplace offering cloud compute at 3-5x lower prices than AWS/GCP. Rent RTX 4090s and H100s by the hour for agent fine-tuning and batch inference. Ideal for cost-sensitive AI workloads that need spot GPU availability.

Serverless GPU infrastructure designed for AI workloads, with simple Python decorators for deployment. Enables agents to invoke fine-tuned models or run heavy inference tasks without managing infrastructure.

Automatic build system that detects your language and creates a Docker image without a Dockerfile. Simplifies containerizing AI agent services that have complex dependency chains. Used by Railway under the hood.

Containerization platform used to package and run agents in isolated environments.

Google Cloud's fully managed platform for building and deploying enterprise AI agents. Includes conversation flows, tool use, RAG, and human escalation. Integrates with Dialogflow, Search, and the broader Google Cloud AI ecosystem. Enterprise SLAs and data residency controls.

Fully managed service for building generative AI agents on AWS. Configure knowledge bases, action groups (Lambda functions as tools), and guardrails through the console or API. Handles orchestration, memory, and tracing without managing infrastructure.

Deploy agent microservices globally close to users with fast cold starts.

Microsoft Azure's enterprise agent service with OpenAI and Azure AI models. Provides tool calling, file search, code interpreter, and persistent threads. Deep Azure Active Directory integration for enterprise auth. SOC 2 and ISO 27001 certified.

Serverless runtime at the edge ideal for lightweight agent endpoints.

Custom LPU (Language Processing Unit) hardware delivering ultra-fast LLM inference — up to 500 tokens/second on Llama and Mixtral models. Critical for latency-sensitive agentic applications where fast token generation directly impacts user experience.

Cloud platform for running open-source LLMs with competitive pricing. Supports fine-tuning, inference, and embeddings. First to offer fast inference on Llama 3 and other frontier open models. Flexible pricing and strong API compatibility with OpenAI's SDK.

Optimized inference platform for open-source models with compound AI and function calling support. FireFunction models are fine-tuned specifically for reliable tool calling in agent pipelines. Compound AI systems enable mixing models for different subtasks.

Run large language models locally with a Docker-like experience. One command to pull and run models like Llama 3, Mistral, and Gemma. OpenAI-compatible REST API makes it a drop-in local replacement. Essential for private, offline, or low-latency agent deployments.

Comprehensive benchmarking environment for evaluating computer-use agents on real desktop tasks. Covers web browsers, file managers, code editors, and productivity apps across Windows, macOS, and Linux — the standard eval suite for GUI-controlling agents.

AI-native browser automation framework built on Playwright. Provides act(), extract(), and observe() commands in natural language alongside traditional Playwright selectors. Combines deterministic automation with LLM intelligence for handling dynamic pages.

Large Action Model framework for natural language web automation. Converts high-level instructions into Selenium or Playwright actions using a specialized action engine, enabling agents to complete complex multi-step web tasks with minimal configuration.

Google's Node.js library for controlling Chrome via the DevTools Protocol. The original headless browser automation library — massive ecosystem of examples and plugins. Lightweight and fast for simple scraping and screenshot tasks agents commonly perform.

Python library making it easy to connect AI agents to the browser. Converts browser state (DOM, screenshot) into LLM-friendly format and translates LLM actions back to browser commands. Supports parallel browser sessions and provides a web UI for session monitoring.

End-to-end browser automation framework with first-class LLM integration patterns. Playwright's reliable cross-browser API makes it the standard for building browser-use agents that need to navigate complex SPAs, handle authentication flows, and extract structured data at scale.

Realistic benchmark for web navigation agents with 812 tasks across realistic web environments (Reddit, GitLab, maps, shopping). Evaluates agents on functional correctness rather than string matching. Standard testbed for browser-use agent research.

Open-source browser infrastructure for AI agents with a developer-first API. Spin up browsers on-demand, manage sessions, handle authentication, and bypass anti-bot systems. Self-hostable Docker image available. Strong focus on reliability for production agent workloads.

Anthropic's API enabling Claude to control a computer by taking screenshots and performing mouse/keyboard actions. Opens the door to agents that can operate any graphical application or website. Available via API with a reference implementation for desktop automation.

Desktop environment for AI agents combining VNC-based screen access with structured action APIs. Agents can click, type, scroll, and observe the screen. Provides both Docker and cloud-hosted environments with Python SDK and replay recording.

Experimental operating system designed from the ground up for AI agent workloads. Explores kernel-level primitives for agent isolation, capability sandboxing, and secure inter-agent communication — an early research project pushing the boundaries of agentic infrastructure.

Browser environment for AI agents with a structured action space instead of raw DOM manipulation. Converts web pages into a graph of perceivable actions (click, fill, select) that agents can take. Reduces complexity of web navigation tasks for language model agents.