Infra AI apps

Sandbox-native cloud for AI agents.

Tensorlake is a cloud platform that gives AI agents their own isolated, stateful sandboxes for running code and calling tools safely. Each sandbox is a microVM that can pause and resume, so long-running agentic loops keep their state across restarts and can run for hours. It exposes serverless Python workflow APIs that scale to zero when idle, and its earlier document-parsing and OCR API now runs as one workload on top. A free tier is available.

Serverless runtime and services for deploying AI agents in production.

AWS's enterprise platform for taking AI agents from prototype to production. A secure, serverless Runtime hosts agents built with any framework (Strands, LangChain, LangGraph, CrewAI) and any model, alongside modular services for Memory, Identity, Gateway (turning APIs/Lambdas into MCP tools), a Code Interpreter, Browser tool, and Observability.

Browsing infrastructure for AI agents — extract, research, automate.

A managed web API that lets agents browse the live web without running headless Chrome. /extract turns pages into Markdown or structured JSON, /generate transforms content on the fly, and /automate clicks, scrolls, and fills forms. Lightweight fetches escalate to full browser automation only when a site needs JS execution.

Open-source observability for logs, metrics, traces and LLMs in one platform.

OpenObserve (O2) is a unified, open-source observability platform that ingests logs, metrics, traces and real-user monitoring with dashboards, alerts and session replay. It adds LLM observability, a natural-language query assistant and an SRE agent for automated incident correlation and root-cause analysis. It runs as a single self-hosted binary or Helm chart, or as fully managed OpenObserve Cloud.

Cloud browser infrastructure for AI agents.

Anchor Browser runs fully managed, stealth Chromium instances in the cloud so AI agents can navigate websites, log in, fill forms, and extract data like a human would. It ships built-in authentication (OmniConnect), an enterprise VPN, a Web Action Cache for deterministic workflows, and scales to many parallel sessions. The Chromium fork is verified by Cloudflare's Web-Bot-Auth to survive bot detection.

Production AI compute platform built by the creators of Ray.

Anyscale is a managed platform for running large-scale AI workloads — distributed training, multimodal data processing, batch inference, and post-training — on Ray, the open-source compute engine. It pools GPUs across AWS, GCP, Azure, and other clouds behind a control plane that adds scaling, observability, and enterprise security (SSO/SAML/SCIM). Built by the UC Berkeley team behind Ray.

AI-native documentation platform that generates and keeps docs current for humans and agents.

Mintlify is a documentation platform that helps teams author, publish, and continuously maintain developer and product docs. It generates and updates content from a codebase, answers reader questions through an embedded AI assistant, and ships an MCP server so AI agents can read the docs directly. Teams get a hosted docs site with a web editor, custom domains, authentication, and analytics.

Open-source MCP gateway, registry, and proxy.

ContextForge is an open-source gateway that sits in front of Model Context Protocol servers, federating them — alongside A2A agents and REST/gRPC APIs — behind a single endpoint with centralized governance. It speaks HTTP, JSON-RPC, WebSocket, SSE, stdio, and streamable HTTP, and ships an admin UI, OpenTelemetry-based observability, Redis-backed caching, and multi-cluster federation. Built and maintained by IBM, it targets production deployments that need to manage many MCP tools at once.

One API for many AI models, plus agent sandboxes and GPU cloud.

Novita AI is an AI and agent cloud for developers that combines serverless model APIs with on-demand compute. A single API serves 120+ text, image, audio, video, and vision models, while Agent Sandbox provides isolated runtimes for tool-using agents and the GPU cloud offers dedicated instances, serverless GPUs, and bare-metal clusters. It advertises sub-50ms time-to-first-token and startup-friendly, usage-based pricing.

Enterprise AI gateway and deployment platform that runs in your own cloud.

A unified platform for deploying, scaling, and governing LLM and agentic AI systems. It pairs an AI gateway that routes and orchestrates calls across providers with infrastructure for hosting models (vLLM, TGI, Triton), fine-tuning, and full-stack observability — deployed inside your own VPC, on-prem, or air-gapped environment with enterprise RBAC and audit logging.

Cloud headless-browser infrastructure for AI agents.

Managed cloud platform that gives AI agents scalable headless browser sessions with built-in stealth mode, proxy rotation, and CAPTCHA solving. Sessions launch in under a second and are driven via API with Playwright, Puppeteer, or Selenium. It also offers HyperAgent, an open-source framework that extends Playwright with natural-language commands, and an official MCP server so LLMs like Claude can browse the live web.

VMs and Git for AI agents — run, version, and deploy the code AI writes.

Freestyle is cloud infrastructure for AI app builders and coding agents. Freestyle VMs are full Linux micro-VMs that provision in under 600ms with live forking and pause/resume, while Freestyle Git provides API-first Git filesystems with branching and GitHub sync that agents can read from and commit to. Together they give an agent a place to write, run, version, preview, and deploy a full-stack app.

Perpetual secure sandboxes that keep AI agents on warm standby.

Blaxel is serverless infrastructure for AI agents: secure sandboxes that spin up in seconds, hibernate after a second of inactivity at zero idle cost, and resume in roughly 25ms with full memory and filesystem state intact. It co-hosts agents, MCP servers, and batch tasks on one backbone, and adds an LLM gateway with caching, rate-limiting, and observability. Snapshots and forking let teams branch agent state for rapid experimentation.

Cloud Devboxes and benchmarks for running AI coding agents at scale.

Runloop provides isolated, cloud-based development environments — "Devboxes" — where AI coding agents safely execute code with full filesystem, build-tool, and compiler access. Teams spin up thousands of Devboxes on demand for large-scale agent tasks and tear them down when done. A benchmarking suite lets you evaluate and compare agents on standardized tests, and enterprise deployments add compliance and VPC options.

Deploy apps, databases, and AI/GPU workloads on any cloud.

Northflank is a developer platform for deploying applications, databases, jobs, and AI/GPU workloads from Git to production. It abstracts away Kubernetes and runs on Northflank's managed cloud or in your own AWS, GCP, Azure, or bare-metal account (BYOC), billed by the second. Teams like Writer, Sentry, and Chai Discovery run on it.

Open-source framework and cloud for realtime voice, video, and physical AI agents.

LiveKit is the realtime infrastructure layer most voice-AI stacks are built on. Its open-source Agents framework wires together streaming speech-to-text, an LLM, and text-to-speech with reliable turn detection, interruption handling, and telephony, so agents can join a session and converse in near real time. You bring your own STT, LLM, and TTS providers; LiveKit handles the low-latency WebRTC transport. It runs self-hosted under Apache 2.0 or as a managed cloud platform.

Serverless GPU infrastructure for real-time AI — voice, video, and LLM workloads.

A serverless GPU platform for deploying real-time AI workloads — voice agents, video models, and LLMs — with cold starts in seconds, instant autoscaling, and multi-region failover. Bring custom code, Dockerfiles, or frameworks like vLLM and pay per second of compute across 12+ GPU types.

Open compute marketplace and RL training stack for agentic models.

Prime Intellect aggregates GPUs from 50+ providers into an on-demand marketplace (1–256 GPUs with SLURM or Kubernetes orchestration) and pairs it with an open reinforcement-learning stack: the prime-rl async training framework, the Verifiers environment library, and a hub of 2,500+ community RL environments. Its INTELLECT model series demonstrates the stack at frontier scale, with weights and training recipes released openly.

Full-stack AI cloud for training and inference at scale.

AI-native cloud built around large NVIDIA GPU clusters — bare-metal and virtualized H100 through Blackwell hardware with InfiniBand networking, managed Slurm and Kubernetes, high-speed storage, and MLOps tooling. Its Token Factory layer adds managed per-token inference for open-weight models. Microsoft signed a five-year capacity deal with Nebius worth up to $19.4B.

Open-access AI cloud: serverless inference + a GPU marketplace.

An AI cloud offering serverless inference for open models (Llama, Qwen, DeepSeek, SDXL, Flux) behind an OpenAI-compatible API, alongside an on-demand GPU marketplace for H100/H200 rentals. It aggregates idle and reserved compute to price inference and GPU hours below centralized clouds. Aimed at developers training, fine-tuning, and serving open-weights models.

Persistent GPU cloud workspaces to build, train, and ship AI.

A cloud platform built around AI Studios — collaborative, persistent GPU workspaces for coding, training models, running inference, and building agents and AI apps. Pay-as-you-go GPUs with a monthly free credit allowance, plus a Pro tier and bring-your-own-cloud for enterprise. Made by the team behind the open-source PyTorch Lightning framework.

On-demand serverless GPU compute for AI, from Python.

A serverless cloud for deploying AI inference endpoints, agent sandboxes, task queues, and containerized GPU workloads with a few lines of Python. It handles fast cold starts, autoscaling, and Docker-in-Docker execution across multiple cloud backends, and supports bring-your-own-compute. The Developer tier is free with recurring monthly credit; paid tiers add team features and scale, billed pay-as-you-go by GPU usage.

Run AI and batch jobs on any cloud or Kubernetes, from one interface.

An open-source framework for running, managing, and scaling AI and batch workloads across Kubernetes, Slurm, and 20+ cloud providers through a single unified interface. It abstracts away per-provider setup, optimizes for cost and GPU availability, and automatically fails over between regions and clouds when capacity is scarce. You run it yourself against your own infrastructure — the software is free and Apache-2.0 licensed; you pay only your own cloud bills.

Browser API for AI agents and apps.

A batteries-included headless-browser sandbox that lets AI agents automate the web without managing browser infrastructure. Steel handles sessions, cookies, proxy rotation, CAPTCHA solving, and page-to-markdown/screenshot/PDF conversion behind a single API, built on Puppeteer and the Chrome DevTools Protocol. Run the open-source image locally via Docker or use Steel Cloud's managed browser fleet.

Secure, elastic sandboxes for running AI-generated code.

Infrastructure for executing AI-generated code in isolated sandboxes — each a full composable computer with a dedicated kernel, filesystem, network stack, and allocated CPU, RAM, and disk. Sandboxes start in under 90ms, snapshot for persistence, and are driven programmatically through SDKs (Python, TypeScript, Ruby, Go, Java), an API, and a CLI. AGPL-3.0 and available as a managed service, self-hosted stack, or hybrid where you bring your own compute.

Secure cloud sandboxes for running AI-generated code.

Open-source infrastructure that spins up isolated cloud sandboxes so AI agents can execute generated code safely. Python and JS/TS SDKs cover code interpretation, data analysis, and computer-use desktops, with per-second billing. Self-host or BYOC/on-prem is supported for enterprise.

Headless browser infrastructure for AI agents.

Managed cloud fleet of headless browsers that let AI agents browse, authenticate, and act on the web at scale. Sessions ship with stealth proxies, automated CAPTCHA solving, and observability, driven via API or the open-source Stagehand framework. Usage-based billing on top of a monthly base plan.

Low-latency inference for open-weights models on custom LPU chips.

GroqCloud serves open-weights models (Llama, DeepSeek, Qwen, Kimi) on Groq's purpose-built LPU hardware, hitting hundreds of tokens per second where GPUs manage tens. OpenAI-compatible API with a free tier; the default when token latency is the product.

Postgres-based backend with auth, storage, and edge functions.

A common default backend when an app needs persistence + auth + realtime. Open-source, self-hostable, very low friction to local dev with the CLI.

Workers + R2 + D1 + Durable Objects. Global edge runtime.

When the app needs to run close to users globally, or when an isolate-per-request model fits the workload better than Node functions. R2 for storage, D1 for SQLite-at-the-edge.

Frontend cloud for React/Next. Edge functions + image opt + analytics.

Next.js-native hosting with fast deploys, edge functions, image optimization, and a free Speed Insights tier. Strong default for the React/Next ecosystem.

Models, datasets, papers, spaces. The AI research commons.

Source-of-truth for open-weights models and datasets. Daily-paper feed for tracking research; Spaces for trying ideas without setting up infra. The default home for any model not behind a paid API.