Hosting & Execution Infrastructure
Where agents actually run — 150+ vendors across 9 categories covering the full hosting and execution landscape.
Where agents actually run — the compute infrastructure for sandboxed, scalable agent execution. This page covers 50+ vendors across 6 tiers, sourced from ClawCamp Research market guides (April 2026).
For LLM inference solutions, see Inference.
Index
- Where to run — Hosting & Execution Platforms · Code Execution Sandboxes · Turnkey Managed Platforms · Agent-Optimized Hosting
- Scale & topology — Agent Orchestration · Cloud Mac Hosting · CI Runners for Agent Iteration · Self-Hosted Infrastructure
- State & control — Memory & Context · Observability & Evaluation · MCP Servers, Registries & Gateways · Identity, Auth & Secrets
- Decision — What Stripe Uses · Sandbox vs. Serverless · Key Trends · Choosing Your Stack · Decision Framework
Agent Hosting & Execution Platforms
The agent hosting market has segmented into six distinct tiers, each with different trade-offs for cost, control, security, and time-to-value.
The Hosting Decision Framework
| Tier | Representative Vendors | Infra Mgmt | Per-Agent Cost | Time-to-Value | Control |
|---|---|---|---|---|---|
| Turnkey / No-Code | ZenClaw, KlausAI, Coral, Lindy | Zero | Highest ($19-400/mo) | Hours | Limited |
| Agent-Optimized | ClawHost, Claw Cloud, Zo Computer | Minimal | Moderate ($1.50-25/mo) | Hours-Days | Moderate |
| Sandbox + Orchestration | E2B, Sprites.dev, Modal, Temporal, LangGraph | Moderate | Usage-based | Days-Weeks | High |
| Serverless | Nebius Serverless, AWS Lambda, Modal | Near-zero | Usage-based | Hours-Days | Moderate-High |
| Cloud Mac | MacStadium, AWS EC2 Mac, Scaleway | Low-Moderate | Monthly subscription | Hours-Days | High |
| Self-Hosted | Hetzner, Contabo, AWS EC2, Railway, Nebius VM | Full | Lowest per-unit | Weeks | Full |
Code Execution Sandboxes
Isolated environments where agents execute generated code safely. This is the single most important layer for autonomous coding agents.
For in-depth coverage, see the dedicated Sandboxes page, which covers:
- Why sandboxes matter for agents
- Core use cases — code execution, tree-of-thought, rollback, persistent environments, multi-tenant fleets, Apple-native, CDEs
- Isolation tier ladder — process → container → gVisor → microVM → VM → bare metal
- All 14 purpose-built agent sandbox vendors
- Contree deep dive — Git-native sandboxing for tree-of-thought agent workflows
- Cloud Dev Environments (CDEs) — Codespaces, Gitpod, Coder, Vercel Sandbox, etc.
- Open-source isolation primitives — Firecracker, Kata, gVisor, etc.
- Agent patterns — checkpoint-explore-commit, golden pool, destructive safety, sandbox-as-context
- Integration examples — MCP, Python SDK, custom harness code
Quick Reference Table
| Vendor | Isolation | Persistence | Cold Start | GPU | Key Strength |
|---|---|---|---|---|---|
| Contree | microVM (Nebius) | Git-like branching | Sub-sec | Yes | Git-style fork/rollback, MCP + Python SDK |
| E2B | Firecracker microVM | Ephemeral / pause | ~150ms | No | Dedicated kernel, SDK-first, SOC 2 |
| Sprites.dev | Firecracker microVM | Hibernate | Instant | No | ~300ms hibernate, zero idle cost |
| Blaxel | Firecracker microVM | Standby + hibernate | ~25ms resume | No | Perpetual sandboxes, scale-to-zero in 1s, SOC 2 / HIPAA / ISO 27001 |
| Daytona | Docker containers | Stateful | ~90ms | Yes | GPU support, fastest creation |
| Modal | gVisor sandbox | Snapshots | Sub-sec | Yes | 50K+ concurrency, full GPU |
| Runloop | Custom hypervisor | Snapshots | Sub-sec | No | 10K+ parallel, SWE-bench focus |
| Northflank | microVM / gVisor | Stateful | Sub-sec | Yes (H100s) | Enterprise VPC, multi-cloud |
| AgentComputer | Ubuntu VMs | 25 GB persistent | Sub-sec | No | Built for Claude/Codex agents |
| + 8 OSS sandboxes | Various | Various | Various | Limited | See full table — includes SmolVM (Apache 2.0 microVM, Mac+Linux) and the Kubernetes agent-sandbox CRD (SIG Apps) |
Turnkey Managed Platforms
Zero infrastructure management — deploy agents in minutes. Best for business teams validating agent ROI without dedicated engineering resources. This category now includes OpenClaw-native platforms, general no-code agent builders, enterprise agent hubs, and autonomous coding agents.
OpenClaw-Native
| Vendor | Isolation | Integrations | Price | Key Strength |
|---|---|---|---|---|
| ZenClaw AI | NVIDIA NemoClaw | Multi-model | $400/mo | 9-second deploy, NVIDIA-backed |
| KlausAI | Isolated cloud | 40+ SaaS tools | $19/mo | Broadest SaaS integrations |
| Coral | Dedicated VM | 500+ integrations | $50/mo | Strongest security, auto cost routing |
| Lindy AI | Managed cloud | 6,000+ integrations | Free/$49.99 | Massive integration catalog |
Enterprise Agent Hubs
Platforms built for enterprise deployments with deep integration into existing enterprise stacks.
| Vendor | Price | Key Strength |
|---|---|---|
| Microsoft Copilot Studio | $200/25K msgs | Deep M365/Teams/Dataverse native integration |
| Google Agentspace | Enterprise paid | Gemini + Google Workspace + enterprise search unified |
| AWS Bedrock Agents | Usage-based (tokens) | Native AWS service/Lambda action group integration |
| Dust | Paid per seat ($29+/user) | Deep workspace data connectors (Notion, Slack, GitHub, Drive) |
| Stack AI | Paid tiered | SOC2/HIPAA compliance focus for regulated industries |
| Sema4.ai | Paid enterprise | Python-based agents with Robocorp RPA lineage |
| Beam AI | Paid enterprise | Vertical agents for ops/finance workflows |
| Orby AI | Enterprise | Learns workflows from user demonstrations |
General No-Code Agent Builders
| Vendor | Price | Key Strength |
|---|---|---|
| Relevance AI | Freemium + usage | "AI workforce" framing with multi-agent teams and tool library |
| n8n AI Agents | Self-host free / Cloud $20+ | Fair-code licensed, 400+ integrations, self-hostable |
| Zapier Agents / Central | Usage-based tasks | 7000+ app integrations out of the box |
| Vellum | Paid tiered | Strong eval/prompt management with production workflows |
| Retool Agents | Paid per user | Bridges agents with internal apps/databases |
| Voiceflow | Freemium + paid | Specialized in voice/chat customer-facing agents |
| Wordware | Freemium | Prompts-as-code editor for agent flows |
| Lutra AI | Paid | Chat-driven multi-step SaaS automation |
| Cognosys | Freemium + usage | Browser-based autonomous web agents for research |
| AgentGPT | Freemium | Simplest "give goal, watch it work" UX |
| MultiOn | Freemium + API | Consumer-grade autonomous web agent |
| SuperAGI | Free OSS + cloud | GUI + marketplace for agent templates and tools |
| Kimi Agent Swarm (kimi.com/agent-swarm) | Consumer + dev platform | Moonshot AI's parallel-task swarm — pre-built verticals for Slides, Websites, Docs, Deep Research, Spreadsheets, and code (Kimi Code) coordinated under one swarm orchestrator. "Scale AI Tasks in Parallel"; developer access via platform.kimi.ai |
Autonomous Coding Agents
Agents specifically focused on software engineering tasks, from PRs to migrations.
| Vendor | Price | Key Strength |
|---|---|---|
| Cognition Labs (Devin) | $500/mo + ACU usage | End-to-end SWE autonomy with VM workspace per task |
| Factory.ai | Paid enterprise | Codebase-aware "Droids" for reviews, migrations, incidents |
| Cursor Background Agents | $20+/mo Pro | Tight IDE coupling with parallel background task execution |
| Replit Agent | $25/mo Core+ | Full app scaffolding + hosting in one workflow |
| Fine.dev | Paid | Focus on autonomous PRs and code tasks |
| Orchestrator.build | BYO Claude API + hosted (early-access) | Spawns many parallel Claude Code agents, each in its own isolated worktree, opens PRs autonomously. Evolved from sirhamy's open-source Phase Golem pipeline runner; useful when you want a Claude-Code-native fleet without building the harness yourself |
| Conductor (conductor.build) | Free during beta | macOS desktop app (closed source) for running Claude Code, Codex, and other agents in parallel. Each agent runs in its own git worktree on your Mac; Conductor handles merge + PR review. Local-first — repo is cloned to your machine, no cloud dependency |
| Superset (superset.sh) | OSS | Editor-style cross-platform alternative to Conductor — runs 10+ parallel Claude Code / Codex agents on a single machine, with worktree isolation, integrated diff viewer, and merge UX |
| Oz (Warp, warp.dev/oz) | Free (4 agents) / Build $18/mo (20 agents) / Max $180/mo (40 agents) + AI & compute usage; self-host for enterprise | Cloud orchestration platform for coding-agent swarms — hundreds of agents in parallel inside Docker-sandboxed cloud environments, multi-repo changes with shared team context, scheduled recurring workflows, CLI / SDK / API, shareable audit trail per run. Launched Feb 10 2026; now the model Warp uses to develop its own open-source codebase |
| Adept | Enterprise | Models trained specifically for UI actions |
Visual Agent IDEs
Open-source and cloud-hosted visual builders for agent workflows.
| Vendor | Price | Key Strength |
|---|---|---|
| AutoGen Studio | Free (OSS) | Microsoft's GUI for AutoGen multi-agent conversations |
| Flowise | Free OSS + cloud | Visual LangChain/LlamaIndex orchestration |
| Langflow | Free OSS + cloud | DataStax-backed LangChain visual IDE |
| CrewAI Enterprise | Free OSS + paid cloud | Role-based crew orchestration paradigm |
| RocketRide (AIDE) | Free OSS + cloud beta | C++ pipeline runtime (3.7K stars) + VS Code extension form factor; MCP-native, TS/Python SDKs, runs on your own infra |
Agent-Optimized Hosting
OpenClaw-specific tooling — agent playgrounds, skills marketplaces, pre-configured integrations — on managed infrastructure. More control than turnkey with lower operational burden than self-hosting.
| Vendor | Isolation | Focus | Price | Key Strength |
|---|---|---|---|---|
| ClawHost | Hetzner VPS | Agent playground, ClawHub | $25/mo | Open-source (MIT), low cost |
| Claw Cloud | Container (Run) | MCP tools | Free/$1.50 | Free tier, 128 vCPU max |
| Zo Computer | Managed server | Personal AI cloud | Free/$18 | Always-on, consumer-oriented |
Agent Orchestration
Durable execution frameworks for running long-lived agent workflows with retry logic, state management, and multi-agent composition. The orchestration layer has become the central battleground for agent infrastructure — with durable execution engines, multi-agent frameworks, and cloud-native workflow platforms all competing.
Durable Execution Platforms
General-purpose workflow engines that handle failure, retries, and state across long-running processes.
| Vendor | Type | Open Source | Price | Key Strength |
|---|---|---|---|---|
| Temporal.io | Durable execution | Yes (MIT) | OSS + Cloud usage | Battle-tested, polyglot SDKs, language-native |
| Inngest | Durable functions for AI | Yes | Free + usage | Step functions with first-class AI agent primitives |
| Trigger.dev | Background jobs with durable runs | Yes | Free + usage | Developer-friendly TS-native with long-running tasks |
| Restate | Durable async runtime | Yes | OSS + Cloud | Lightweight single-binary durable execution |
| DBOS | Durable execution in Postgres | Yes | OSS + Cloud | Stores state in your Postgres, no separate service |
| Hatchet | Distributed task queue + workflows | Yes | OSS + Cloud | Postgres-backed, Temporal-lite ergonomics |
| Windmill | OSS dev platform for workflows | Yes | OSS + Cloud | Scripts + flows + UIs in one self-hostable platform |
| Orkes (Conductor) | Managed Netflix Conductor | Yes | OSS + Cloud | Proven at Netflix scale, microservices orchestration |
| Uber Cadence | Temporal's predecessor | Yes | OSS | Uber-backed durable workflow engine |
| Resonate HQ | Distributed async/await | Yes | OSS | Durable promises as core primitive |
| Kestra | Declarative YAML orchestrator | Yes | OSS + cloud | YAML-first, language-agnostic |
| Convex Workflows | Durable workflows in reactive DB | No | Free + usage | Tightly integrated with reactive DB backend |
| Cloudflare Workflows | Durable execution on Workers | No | Usage-based | Edge-native durable execution |
| Upstash Workflow | Serverless durable workflows | No | Free + usage | QStash-backed, serverless-first |
| Vercel Workflow SDK | Durable workflow execution for AI agents | No | Usage-based | Streaming + cancellation + reconnectable streams, powers Vercel Open Agents, integrates with AI SDK |
Inngest AgentKit
Inngest's primary differentiator vs. Temporal / Restate / DBOS is the dedicated agent SDK layered on top of the durable-execution core. AgentKit (TS, MIT) ships Agent / Network / Router / State primitives directly — every tool call, every model inference, every network hop is a durable step under the hood, so retries, observability, human-in-the-loop pauses, and replay come for free. step.ai.wrap and step.ai.infer let any AI SDK call inherit that durability without rewriting it. The TS-native ergonomics, Vercel/Cloudflare-friendly deployment story, and "agents as graphs of durable functions" framing make it the obvious bridge between the Agent-Specific Orchestration Frameworks row below and the Durable Execution Platforms row above.
Temporal for agents
Temporal (temporalio/temporal, MIT, 20K stars) is the language-native end of this market — workflows are ordinary Python / TS / Go / Java functions, and Temporal makes them survive crashes, restarts, and 30-day sleeps. For agents that's exactly the shape of the problem: a tool call is an activity (auto-retried, idempotent), an LLM turn is an activity (with timeouts), the conversation is a workflow (event-history-replayable). Teams running Stripe Minions-style fleets at scale increasingly land on Temporal because the same engine that orchestrates the agent loop also orchestrates the payments pipeline next door. Battle-tested at Uber / Snap / Coinbase / Datadog scale — the trade-off vs Inngest is operational weight (worker cluster + history service) in exchange for polyglot SDKs and a 7+ year production track record.
Cloud Provider Workflow Engines
Managed state-machine services from the major clouds.
| Vendor | Price | Key Strength |
|---|---|---|
| AWS Step Functions | Per-transition | Deep AWS integration, visual state machines |
| Azure Durable Functions | Consumption | Orchestrator pattern in Azure Functions |
| Google Cloud Workflows | Per-step | Serverless integration orchestration on GCP |
Agent-Specific Orchestration Frameworks
Open-source and managed frameworks purpose-built for multi-agent systems.
| Vendor | Type | Price | Key Strength |
|---|---|---|---|
| LangGraph / LangGraph Platform | Stateful graph agents | OSS + Cloud ($39/user) | Graph-based agent state machines + create_agent factory (LangChain 1.0 default since Oct 2025; supersedes deprecated create_react_agent), LangSmith integration. Reference open harness: Deep Agents |
| CrewAI | Multi-agent role orchestration | OSS + enterprise | Role/task/crew abstraction |
| Microsoft Agent Framework | Unified successor to AutoGen + Semantic Kernel | OSS (MIT, 1.0 RC Feb 2026) | Merged AutoGen + Semantic Kernel into one framework with workflow + actor model; Python and .NET; Azure-aligned but cloud-neutral |
| Microsoft AutoGen | Multi-agent conversation framework | OSS (maintenance — superseded by Agent Framework) | Conversational multi-agent patterns |
| Paperclip | Hierarchical multi-agent OSS framework — CEO / manager / worker org chart | OSS (44.9K stars in 3 weeks, March 2026) | Doesn't build agents — orchestrates existing ones (Claude Code, OpenClaw, Codex, scripts, webhooks) into "AI companies" with budgets, reporting lines, audit trails |
| OpenAI Agents SDK | Handoff-based agent orchestration (production successor to the deprecated Swarm educational library) | OSS (MIT, 27K Python / 3.1K JS) | First-party OpenAI framework — agent / runner / handoff primitives, tracing, guardrails, sessions, MCP, sandbox agents (v0.14+); provider-neutral via the Responses API. See Approaches: OpenAI Agents SDK |
| Google ADK | Model-driven agent framework | OSS (Apache 2.0, 20K stars) | Google's first-party framework — same agent/tool/session shape as OpenAI Agents SDK + Strands, deep Gemini + Vertex AI integration, Agent Engine / Cloud Run deployment templates; unusually polished samples repo |
| Strands Agents | Model-driven agent SDK | OSS (Apache 2.0, 5.9K stars) | AWS-incubated answer to OpenAI Agents SDK / ADK — 13+ model providers, native MCP, multi-agent composition, experimental bidirectional voice streaming; first-class Lambda / Fargate / Bedrock AgentCore deployment |
| LlamaIndex Workflows | Event-driven agent workflows | OSS + cloud | Event-driven steps tied to LlamaIndex RAG |
| Pydantic AI | Type-safe agent framework | OSS | Pydantic-grade type safety for agents |
| Burr | State-machine agent framework | OSS | Explicit state machine + telemetry for agents |
| Haystack Agents | deepset's agent framework | OSS + cloud | Pipeline-oriented RAG + agents |
| Mastra | TypeScript AI agent framework | OSS | TS-native agents with workflows + evals |
| Kagent | K8s framework | Free OSS | K8s-native, CNCF, multi-framework |
Data & ML Orchestrators (Cross-Over)
Traditional data/ML orchestrators increasingly used for AI agent workflows.
| Vendor | Price | Key Strength |
|---|---|---|
| Prefect | OSS + Cloud | Pythonic DAGs with dynamic runtime flows |
| Dagster | OSS + Cloud | Asset-first model, strong for data+AI pipelines |
| Apache Airflow / Astronomer | OSS + managed | Industry-standard data orchestration |
| Flyte / Union.ai | OSS + Cloud | Typed, reproducible ML pipelines, K8s-native |
Cloud Mac Hosting
Dedicated macOS environments for agents needing Apple-native capabilities — iMessage, Xcode, iOS Simulator, Neural Engine inference. A hard requirement for agents in the Apple ecosystem.
Dedicated Mac Hosting
| Vendor | Hardware | Isolation | Price | Key Strength |
|---|---|---|---|---|
| MacStadium | M1/M2/M4 Mac Mini + Max | Orka virtualization | Monthly | Largest Mac cloud, <1s VM launch |
| AWS EC2 Mac | M4/M4 Pro/Max (Nitro) | Bare metal | Hourly (24hr min dedicated) | Full AWS VPC/EBS integration |
| Scaleway Mac | M4 Mac Mini | Bare metal | Hourly | EU sovereign hosting (Paris DC) |
| MacinCloud | M1/M2 | Managed / dedicated | Monthly | Global presence, managed CI/CD |
| Roundfleet | M4/M4 Pro/Max Mac Mini | Dedicated | Monthly | High availability, fast provisioning |
| Flow Swiss | M-series Apple silicon | Bare metal | Monthly | Swiss data sovereignty |
| MacinCloud | M1/M2 Macs | Managed | Monthly | Global presence, turnkey setup |
| HostMyApple | M1/M2 Macs | VPS + dedicated | Monthly | Budget M-series rentals, 3 DCs |
| Macly | M4 Mac Mini | Dedicated | Monthly | 24/7 support included, no sales calls |
| Mac-in-a-Box | Dedicated Mac hardware | Bare metal | Monthly flat | Budget dedicated Mac rentals |
| Nimble | Apple silicon | Dedicated | Monthly | Dedicated Apple silicon hosts |
| MacWeb | Mac hosting | Dedicated | Monthly | Long-running Mac hosting |
| Oakhost | Apple silicon | Dedicated | Monthly | EU-hosted Apple silicon |
Mac CI Runners
Managed CI services with macOS runners — useful for build-and-test agent workflows that don't need persistent Mac state.
| Vendor | Price | Key Strength |
|---|---|---|
| Apple Xcode Cloud | Usage (compute hours) | Native Apple integration, App Store signing |
| GitHub Actions macOS Runners | Per-minute | Integrated with GitHub workflows |
| CircleCI macOS | Per-credit | Mature iOS CI pipelines |
| Codemagic | Free tier + usage | Flutter/mobile-app specialized |
| Bitrise | Per-seat + usage | Mobile-app workflow library |
| Appcircle | Paid tiered | Mobile-only DevOps platform |
| Cirrus CI Mac | Usage-based | Generous free tier for OSS, per-minute Apple silicon |
CI Runners for Agent Iteration
When agents author PRs autonomously, the CI run is the feedback loop. The default GitHub-hosted runners were sized for human PR cadence; agent fleets pushing dozens of PRs an hour expose them as the bottleneck. A new generation of drop-in runner replacements has emerged specifically pitching "faster CI = faster agent iteration" — same runs-on: swap, same workflows, 2–10× wall-clock improvements on typical jobs. (For macOS-specific runners, see the Mac CI Runners table above.)
| Vendor | Hardware | Isolation | Speed vs. GitHub-hosted | Pricing | Agent Angle |
|---|---|---|---|---|---|
| Blacksmith (blacksmith.sh) | Bare-metal gaming CPUs (top single-core perf), persistent NVMe layer cache | Bare metal | Up to 2× faster; up to 40× faster Docker builds via warm caches | ~$0.004/min for 2 vCPU x64 (half of GitHub-hosted); 3,000 free min/mo; macOS M4 $0.08/min | YC W24, $10M Series A (Sep 2025, GV-led). Launch headline was literally "Unblock AI Development with Fast CI." 800+ paying customers including Vercel, Supabase, Clerk, Mercury, Ashby. ARM + macOS M4 + Windows Server 2025 (beta). |
| Tenki Runners (tenki.cloud/products/runners) | x64 Linux + Apple Silicon M4 Pro | Firecracker microVM, destroyed per job | "30% faster, up to 60–90% cheaper" | $0.002/core-min x64, $0.080/core-min macOS; Starter $10/mo credits, Team $200/mo | From Luxor Technology (Seattle). Sister product to the Tenki Sandbox for AI-agent code execution — same operator across the agent's "run untrusted code" and "verify the PR" surfaces. SOC 2 Type II. |
| Depot (depot.dev) | Bare-metal x64 + ARM, persistent build cache | Container / VM | 2–10× faster Docker / Bazel builds; faster Actions runners | Per-minute, free tier for OSS | Started as a Docker build accelerator; runners are the natural extension. Cache-first architecture maps well to agent loops that rebuild the same project hundreds of times. |
| Namespace (namespace.so) | Bare-metal x64 + ARM | microVM | 2–4× faster, similar cost as GitHub-hosted | Per-minute, generous free tier | Sells both fast Actions runners and remote dev environments — closest competitor to the Ona / Daytona category on the dev-env side and Blacksmith / Tenki on the CI side. |
| BuildJet (buildjet.com) | Bare-metal x64 + ARM | VM | ~2× faster | ~30% cheaper than GitHub-hosted | One of the earliest faster-runner shops; mature, no AI-specific positioning — still the conservative pick. |
| RunsOn (runs-on.com) | Your AWS account, any EC2 SKU | EC2 instance | Up to 10× faster on right-sized hardware | $0 per minute — you pay AWS only; flat license fee | Self-hosted: the runners live in your VPC, so secrets, GPUs, and private network access are all native. Right fit when an agent fleet needs the same VPC posture as production. |
| Ubicloud (ubicloud.com) | Bare-metal | VM | ~2× faster | Up to 10× cheaper than GitHub-hosted | Open-source, self-hostable; positioned as "open AWS." Same drop-in runs-on: swap with the cost discipline of running on Hetzner-class infra. |
Why it matters for agents. Blacksmith CEO JP Jayaprakash's framing on the Series A: "This is even more important for teams that are using AI codegen tools and want to move quickly." If a Copilot Coding Agent or a fleet of Claude managed agents opens 30 PRs an hour and each waits 8 minutes for CI to turn green before merge-or-iterate, the harness is throttled by the runner queue, not the model. Cutting CI from 8 minutes to 90 seconds is often a more leveraged investment than upgrading the model. Compare to the parallel logic in Sandboxes — Tier 3 latent opportunities: the faster the verification signal, the more iteration the harness can afford.
Choosing. Default to Blacksmith if you want the lowest-touch SaaS migration with explicit AI-codegen positioning. Pick Tenki Runners if you also need an agent code-execution sandbox from the same vendor (the Runners + Sandbox combo). Pick Depot if your bottleneck is Docker / Bazel builds rather than the runners themselves. Pick RunsOn or Ubicloud if you want the runners inside your own AWS / Hetzner footprint. Namespace is the right call if you also want remote dev environments from the same provider.
Self-Hosted Infrastructure
Full control over the entire stack. Lowest per-unit cost but highest operational burden. Split across three sub-categories: specialized GPU clouds (for model serving), general-purpose clouds, and VPS/bare-metal providers.
Specialized GPU Clouds
Purpose-built GPU infrastructure for AI workloads — the most cost-effective path for running open-weight models at scale.
| Vendor | GPU Inventory | Price Model | Key Strength |
|---|---|---|---|
| Nebius | B300/B200/H200/H100/L40S | Hourly + preemptible | AI-native cloud, Aether 3.5 serverless |
| CoreWeave | H100/H200/GH200 at scale | Hourly reserved | Purpose-built H100 cloud, largest dedicated GPU fleet |
| Lambda Labs | H100/GH200 | Hourly | On-demand for ML researchers |
| RunPod | Community + datacenter GPUs | Per-second | Community cloud + serverless GPU endpoints |
| Paperspace (DigitalOcean) | H100/A100 + notebooks | Hourly | Gradient ML notebooks included |
| Genesis Cloud | H100/A100 (EU) | Hourly | 100% renewable energy, EU-based |
| Voltage Park | H100 at non-profit rates | Hourly reserved | Non-profit H100 capacity |
| TensorDock | Mixed GPU marketplace | Hourly | Cheap community GPU rentals |
| Crusoe | Flared-gas-powered GPU | Contract | Climate-aligned data centers |
| FluidStack | Aggregated GPU supply | Hourly | GPU aggregator with competitive prices |
| Vast.ai | Decentralized GPU marketplace | Bid-based | Cheapest consumer GPU spot market |
| Together AI | Training clusters | Hourly + tokens | Training + fast inference unified |
| Cudo Compute | Distributed GPU cloud | Hourly | Distributed GPU availability |
| Hyperstack (NexGen) | H100/H200 reserved | Hourly | NVIDIA-partner GPU cloud |
General-Purpose Cloud Platforms
| Vendor | Type | GPU | Starting Price | Key Strength |
|---|---|---|---|---|
| AWS EC2 | Virtual machines | Yes | On-demand | Full control, broadest GPU range |
| AWS Lambda | Serverless functions | No | $0.20/1M req | Scale-to-zero, vast ecosystem |
| Google Cloud Compute | VMs + TPU | Yes | Per-second | TPU exclusivity, deep BigQuery integration |
| Azure VMs | Virtual machines | Yes | Per-minute | Enterprise + OpenAI partnership |
| Oracle Cloud (OCI) | Full-stack cloud | Yes | Generous free tier | Always-free ARM Ampere instances |
| IBM Cloud | Enterprise + bare metal | Yes | Varies | Enterprise/regulated industry focus |
| Alibaba Cloud | Largest China cloud | Yes | Pay-as-you-go | Dominant APAC/China presence |
| Tencent Cloud | China cloud + gaming | Yes | Pay-as-you-go | Gaming/media APAC specialization |
| DigitalOcean | Droplets (VMs) | Yes | $4/mo | Developer-friendly, managed K8s |
| Linode (Akamai) | VPS + edge cloud | Yes | Hourly/monthly | Simple pricing, Akamai edge network |
| Vultr | Global VPS + GPU | Yes | Hourly | 30+ global locations, bare metal options |
| Scaleway | EU-sovereign cloud | Yes | Hourly | EU data sovereignty, ARM options |
| UpCloud | High-performance VPS (EU) | No | Hourly | MaxIOPS storage performance |
| Fly.io | Firecracker VMs | Ltd | Per-second | 30+ regions, Sprites integration |
| Render | Persistent containers | No | Free / $25 | SOC 2, HIPAA, ISO 27001 |
| Railway | Containers | No | $5/mo | Hard spending caps, MCP server |
| Heroku | Dynos (containers) | No | $5/mo | Simple git-push, add-ons |
| Exoscale | Swiss/EU cloud | No | Hourly | Swiss data sovereignty |
| CloudSigma | EU/US VPS | No | Hourly | Fully customizable resource sliders |
VPS & Bare Metal Providers
Budget-focused providers best suited for self-hosting agent infrastructure at lowest per-unit cost. See the VPS for agents deep dive below for type definitions, pricing detail, and decision guidance.
| Vendor | GPU | Starting Price | Key Strength |
|---|---|---|---|
| Hetzner | Yes | EUR 3.79 | Exceptional price-to-performance |
| OVHcloud | Yes | EUR 3.50 | EU data sovereignty, 40+ DCs |
| Hostinger | No | $4.99/mo | Budget-friendly, global reach |
| GTHost | Yes | Custom | AI/ML optimized dedicated servers |
| Contabo | Yes | EUR 3.60 | Aggressive pricing, H200 available |
| Kamatera | No | Hourly | Highly configurable VMs |
| Latitude.sh | No | Hourly/monthly | Bare metal in 20+ regions |
| Equinix Metal | No | Hourly | Bare metal at IX colos, premium interconnect |
| Leaseweb | Yes | Monthly | Large dedicated server inventory |
| phoenixNAP | Yes | Monthly | Enterprise bare metal provider |
VPS for Agents
A Virtual Private Server (VPS) is a virtual machine with its own dedicated slice of CPU, RAM, storage, and IP running on shared physical hardware. You get root access, choose your OS (usually Linux, sometimes Windows), and run whatever you want — from a single long-lived process to a full Docker stack hosting multiple agents.
For agents, VPS sits in an interesting spot on the hosting ladder: cheaper and simpler than a bare-metal server, more persistent and unconstrained than a sandbox or a serverless function, less locked-in than a managed agent platform. Much of the infrastructure powering hobbyist and indie agent deployments — private Claude Code runners, n8n automation servers, always-on harnesses, personal LLM gateways — runs on a single $5-a-month VPS.
Why VPS fits agent workloads
- Root access and arbitrary runtimes. Install any CLI (Claude Code, Codex, Goose, Aider, the full Docking Station set), any language runtime, any database, any VPN client. No platform-imposed restrictions on what the agent can spawn.
- Persistent state across runs. Unlike ephemeral sandboxes, a VPS keeps files, caches, cloned repos, and credentials between sessions. Good for iterative agent loops that benefit from a warm workspace (populated node_modules, pre-indexed codebase, warm model cache).
- No cold starts, no timeouts. Long-running background workers — durable agent schedulers, queue consumers, MCP servers, scraping pipelines — run indefinitely. Serverless platforms kill after 5-15 minutes; a VPS runs for months.
- Lower isolation cost than microVMs. Firecracker-per-task (E2B, Sprites, Contree) has ~150ms boot and per-second billing; a $5/mo VPS is essentially free-per-invocation once you own it. For trusted agents running your own code, the hypervisor-level isolation of a dedicated microVM is overkill.
- Custom networking. Static IP, open ports, WireGuard / Tailscale mesh for multi-agent coordination, reverse tunnels into home labs. Sandboxes typically restrict inbound networking; a VPS does not.
- Deterministic cost. Flat monthly fee. No surprise bills from a runaway agent loop — worst case the VPS's CPU pegs at 100%, not your credit card.
The classic pattern: one VPS runs the orchestrator (durable workflow engine, MCP gateway, scheduled-task runner), and when the agent needs to execute untrusted or destructive code, it delegates to a dedicated sandbox (E2B, Sprites, Contree). VPS is the always-on control plane; sandboxes are the ephemeral execution plane.
The three VPS flavors
| Type | What you manage | What the provider manages | Best for |
|---|---|---|---|
| Unmanaged / Self-managed | OS patches, security hardening, backups, monitoring, app stack | Hypervisor, network, hardware | Experienced operators; cheapest tier; full control |
| Managed | Your application and data | OS updates, security patches, often backups and monitoring too | Teams that want a VPS without sysadmin burden; typically 2-3x the price |
| Cloud VPS | Your app + optional scaling config | Hypervisor, network, elastic resources, often snapshots and load balancing | Agents with variable load; pay-as-you-go scaling without rearchitecting |
Unmanaged is the default for agent hobbyists and the Docking Station-style self-hosted stack — you're already comfortable in a shell, and the savings compound. Managed pays for itself once the operational toil exceeds the price delta (usually true for small businesses without an ops person). Cloud VPS (DigitalOcean Droplets, Linode, Vultr, Lightsail) is the middle ground: hourly billing, snapshot-based backups, easy resize — closer to a cloud VM but priced and packaged like a VPS.
Provider comparison (entry tier)
Prices are starting monthly prices for the lowest published tier; availability of promotional pricing varies by region and commitment length. Always check current pricing before committing.
| Provider | Entry price | Type | Agent-relevant notes |
|---|---|---|---|
| IONOS | ~$2/mo | Cloud VPS | Cheapest mainstream entry tier; EU/US DCs; good for always-on control planes and webhooks |
| Hostinger VPS | ~$6.49/mo | Managed / Unmanaged | AI Assistant + Docker templates, 1-click LLM stacks, good for non-sysadmin users |
| DigitalOcean Droplet | $4/mo | Cloud VPS | Best developer experience, 1-click marketplace apps (Ollama, n8n, Langfuse), managed K8s nearby for scale-out |
| OVHcloud VPS-1 | ~$6.46/mo | Cloud VPS | EU data sovereignty, 40+ DCs, optional GPU tiers higher up the stack |
| Amazon Lightsail | From ~$3.50/mo | Cloud VPS | Fixed-price AWS on-ramp; simplest path to layering in S3, SES, Route53 around the VPS |
| Contabo | EUR 3.60 (~$4) | Cloud VPS | Aggressive RAM/storage per dollar; popular for self-hosting agent inference and vector DBs |
| Hetzner CX | EUR 3.79 (~$4) | Cloud VPS | Exceptional price-to-performance in EU; dedicated-vCPU tiers ideal for model-adjacent workloads |
| Linode (Akamai) | $5/mo | Cloud VPS | Predictable pricing, Akamai edge network, GPU plans for inference |
| Vultr | $2.50/mo | Cloud VPS | 30+ global regions, bare-metal and GPU plans in the same console |
When a VPS is the right fit
Choose a VPS when you want:
- An always-on agent control plane — durable workflow runner (Temporal worker, n8n, Trigger.dev self-hosted), MCP gateway, scheduled-task loop, webhook receiver.
- A personal self-hosted assistant stack — OpenClaw / Letta / a CLI harness plus a local vector DB plus a model gateway, all in one place.
- A shared dev target for agents — devbox-style box where agents SSH in, run tests, and leave artifacts behind between runs.
- A private VPN / Tailscale exit node to give agents access to home-lab resources or region-locked services.
- A cheap, always-on hobbyist deployment — the kind of workload that would cost $50+/mo on serverless but costs $4/mo here.
Choose something else when:
- You need Firecracker-level isolation per task (untrusted LLM code, multi-tenant agent runs) → use E2B, Sprites.dev, Contree; see Sandboxes.
- You need horizontal autoscaling to hundreds of concurrent agents → use serverless (Nebius, Modal, AWS Lambda) or a managed sandbox platform.
- You need GPU inference at scale → see Inference (Nebius, Together, Fireworks, Groq) rather than trying to run vLLM on a single VPS.
- You need managed compliance (SOC 2 / HIPAA) without rolling it yourself → Render, Fly.io, or a hyperscaler will get you further than a raw VPS.
How VPS slots into the agentic-engineering stack
Think of it as the persistent substrate underneath the ephemeral sandbox layer:
User / CI trigger
│
▼
VPS (always on) ← orchestrator, scheduler, MCP gateway, harness
│
├─────► Sandbox (per task, ephemeral) ← untrusted exec, test runs
│
├─────► Inference API / Platform ← LLM calls
│
└─────► Object storage / DB (persistent) ← artifacts, memory, traces
Agents that live on a VPS can still reach into the entire rest of the stack — they just do so from a stable, cheap, fully-owned home base instead of being reborn from scratch on every invocation.
Agent Memory & Context Infrastructure
Stateless agents forget everything between sessions. Memory infrastructure gives agents persistent, retrievable context across conversations, users, and sessions — transforming them from stateless tools into adaptive systems that learn and improve.
Purpose-Built Agent Memory
Memory layers designed specifically for AI agents, with multi-level scoping (user, session, agent) and semantic retrieval.
| Vendor | Type | Price | Key Strength |
|---|---|---|---|
| Mem0 | Agent memory layer | OSS + cloud | Self-improving memory, 26% accuracy gain over OpenAI Memory |
| Letta (formerly MemGPT) | Stateful agent server | OSS (Apache 2.0, 23K stars) + cloud | Reference memory-first agent framework. Runs as a server; agents are durable, REST-addressable resources with a four-block memory hierarchy (core / archival / recall / message buffer) — the MemGPT paper's design productized. Ships Letta Code CLI, model-agnostic. The default pick when you want OS-level memory without GBrain's repo-as-source-of-truth opinion |
| Zep | Long-term memory + knowledge graph | OSS + cloud | Temporal knowledge graph for agents |
| Cognee | AI memory engine | OSS + cloud | Knowledge graph + vector hybrid memory |
| Graphlit | Knowledge API for agents | Usage-based | RAG + knowledge graph as a service |
| Motorhead | Chat memory server | OSS | Lightweight chat history service |
| Basic Memory / OpenMemory | OSS agent memory protocols | Free OSS | Standardized memory protocols |
| MemMachine | Universal memory layer for agents | OSS | Persistent multi-session memory that works across models and environments — drop-in alternative when you don't want to lock into Mem0 / Letta APIs |
| GBrain | Self-wiring knowledge graph + memory + durable job queue | OSS (MIT, 19K stars) | Garry Tan's production memory layer — markdown as system of record, Postgres + pgvector engine (PGLite or Supabase), typed-edge graph extracted with zero LLM calls, 29 built-in skills, "Minions" Postgres-native job queue (753 ms vs 10 s + sub-agent spawn); pairs with GStack |
Vector Databases
The underlying infrastructure for semantic search and retrieval over agent memory.
| Vendor | Type | Price | Key Strength |
|---|---|---|---|
| Pinecone | Managed vector DB | Usage-based | Serverless pioneer, mature vector index |
| Weaviate | Vector DB with modules | OSS + cloud | Hybrid search + built-in modules |
| Chroma | Embedding DB for AI | OSS + cloud | Dev-friendly, simple API |
| Qdrant | Rust-based vector DB | OSS + cloud | Rust performance, rich filtering |
| Milvus / Zilliz Cloud | Scalable vector DB | OSS + cloud | Billion-scale vector workloads |
| LanceDB | Embedded vector DB | OSS + cloud | Serverless embedded + multimodal |
| Turbopuffer | Object-store-backed vectors | Usage-based | Cheap vector search on S3 |
| MongoDB Atlas Vector Search | Vectors in MongoDB | Cluster-based | Unified document + vector store |
| pgvector (Neon, Supabase) | Postgres vector extension | OSS / DB-tier | Vectors alongside relational data |
| Redis / Redis Stack | In-memory vector + cache | OSS + cloud | Lowest-latency vector + KV |
| Marqo | End-to-end vector search | OSS + cloud | Multimodal embeddings built-in |
| Typesense | Open-source search + vector | OSS + cloud | Typo-tolerant hybrid search |
| Elasticsearch / OpenSearch | Search + kNN | OSS + cloud | Mature search with vector support |
| Vespa | Full search + vector platform | OSS + cloud | Yahoo-scale search + vector hybrid |
| Vectara | Managed RAG-as-a-service | Usage-based | End-to-end RAG with hallucination scoring |
| Azure AI Search | Managed hybrid search | Usage-based | Deep Azure/OpenAI integration |
| Vertex AI Vector Search | Google's ScaNN service | Usage-based | Google's internal ScaNN algorithm |
| SingleStore | HTAP with vectors | Paid | Transactional + analytical + vector |
Graph Databases for Agent Knowledge
For agents reasoning over structured knowledge rather than flat key-value pairs.
| Vendor | Price | Key Strength |
|---|---|---|
| Neo4j + GraphRAG | OSS + cloud | Industry-standard graph DB, GraphRAG-native |
| Mem0g (Graph Memory) | OSS + cloud | Mem0's graph feature mapping entity relationships |
Search APIs for Agents
Memory is what an agent remembers; search APIs are what it can find out fresh. The market has converged on a small number of LLM-tuned search/extract endpoints that return clean markdown rather than raw HTML.
| Vendor | Endpoints | Price | Key Strength |
|---|---|---|---|
| Tavily | Search · Extract · Crawl · Map · Research | Free tier + usage | LLM-native search built for agents. The five-endpoint surface (released through 2025–26) covers ad-hoc lookup (Search), URL → markdown (Extract), domain crawl (Crawl), sitemap-style enumeration (Map), and multi-step deep-research with structured output (Research). MCP server is a default install in Claude Code, OpenCode, and the Anthropic Marketplace. Pairs cleanly with Mem0 / Letta — Tavily fetches, Mem0 / Letta remember |
| Firecrawl | Scrape · Crawl · Map · Extract · Deep Research | OSS (AGPL) + cloud | 124K-star OSS alternative; same shape, self-hostable, JS/TS-native |
| Exa | Neural search · Contents · Find Similar · Answer | Usage-based | Neural-embedding search, strong on niche / academic queries where keyword search degrades |
Agent Observability & Evaluation
As agents move to production, observability becomes critical. This category splits into tracing (what happened), evaluation (was it correct), and guardrails (prevent bad outcomes).
LLM & Agent Tracing / Observability
Platforms for tracing agent runs, debugging failures, and monitoring costs in production.
| Vendor | Type | Price | Key Strength |
|---|---|---|---|
| LangSmith | LangChain-native tracing/evals | Free tier + usage | Deepest LangChain/LangGraph integration |
| Langfuse | OSS LLM observability | OSS + cloud | Self-hostable LangSmith alternative; acquired by ClickHouse Jan 2026 |
| OpenObserve | Unified observability (logs / metrics / traces / RUM + LLM) | OSS (AGPL-3.0) + cloud | 3.0 (April 2026) ships an autonomous AI SRE agent for incident response; one platform for infra and LLM monitoring rather than a dedicated LLM tool |
| Arize Phoenix / Arize AX | OSS + enterprise LLM ops | OSS + cloud | OTel-based, strong ML+LLM crossover |
| Weights & Biases Weave | LLM tracing on W&B | Paid tiered | Integrated with W&B ML experiment tracking |
| Helicone | LLM proxy + observability | OSS + cloud | Simple one-line proxy onboarding |
| Datadog LLM Observability | LLM tracing in Datadog | Per-span pricing | Unified with existing APM |
| Dynatrace AI Observability | Auto-instrumented AI ops | Enterprise | Auto-discovery AI pipelines |
| Honeycomb for LLMs | High-cardinality LLM tracing | Usage-based | BubbleUp for LLM anomaly detection |
| OpenLLMetry / Traceloop | OTel-standard LLM semantics | OSS + cloud | Vendor-neutral OTel for LLMs |
| AgentOps | Agent session replay + costs | Free + paid | Session-replay view of agent runs |
| PromptLayer | Prompt versioning + logs | Free + paid | Prompt-registry-first workflow |
| LangWatch | LLM monitoring + evals | OSS + cloud | European LLM observability |
| Lunary | OSS LLM observability | OSS + cloud | Lightweight open alternative |
| Literal AI | Eval + observability (Chainlit) | Paid | Chainlit-native LLM ops |
| Fiddler AI | ML + LLM monitoring | Enterprise | Enterprise explainability + drift |
Evaluation & Testing
Tools focused on measuring agent quality and regression testing.
| Vendor | Type | Price | Key Strength |
|---|---|---|---|
| Braintrust | Eval-first LLM platform | Paid tiered | Eval-driven prompt iteration |
| Patronus AI | Automated LLM eval | Paid | Pre-built eval models (Lynx, etc.) |
| Galileo | LLM eval + guardrails | Paid | ChainPoll metrics for hallucinations |
| HumanLoop | Prompt + eval management | Paid | PM/engineer collaborative prompt eng |
| Comet Opik | OSS LLM eval | OSS + cloud | Comet ML lineage, strong evals |
| TruLens | Eval framework (Snowflake) | OSS | Groundedness/answer/context triad |
| Ragas | OSS RAG eval framework | OSS | De-facto RAG metrics library |
| DeepEval | Unit-test-style LLM evals | OSS | Pytest-style LLM assertions |
| Confident AI | DeepEval hosted platform | Paid | Hosted DeepEval with regression testing |
| Future AGI | End-to-end eval + tracing + simulations + guardrails + gateway | OSS (Apache 2.0, self-hostable) | One platform for shipping self-improving agents — covers eval through optimization rather than picking one slice |
| Anthropic Bloom | Automated behavioral evals for frontier models | OSS (Anthropic Research) | Quantifies frequency + severity of researcher-specified behaviors across auto-generated scenarios; built for safety / red-team eval rather than task accuracy |
Guardrails & Safety
Runtime safeguards preventing agents from producing harmful or policy-violating outputs.
| Vendor | Type | Price | Key Strength |
|---|---|---|---|
| NVIDIA NeMo Guardrails | Programmable guardrails | Free OSS | Colang DSL for conversation rails |
| Guardrails AI | Validation library | OSS + cloud | Structured output + validators library |
| Lakera | LLM security guardrails | Paid | Prompt injection defense specialist |
| Protect AI | AI security posture | Enterprise | MLSecOps / AI red-team toolkit |
| WhyLabs / LangKit | LLM safety monitoring | Free + paid | Safety/guardrails metrics focus |
| LlamaFirewall (Meta) | Agent-specific guardrails | Free OSS | Final defense layer against prompt injection, agent misalignment, insecure code; designed to wrap agent harnesses, not just chat |
| LLM Guard | Drop-in input/output scanner | Free OSS | 35 scanners for prompt injection, PII leaks, toxicity; pip-install-and-go |
MCP Servers, Registries & Gateways
Once an agent harness picks up MCP support, the next decision is where the MCP servers actually run — local subprocess, hosted remote, or behind a gateway that fans out to dozens. The market split into four flavors during late 2025 / early 2026:
- Registries — discover and install community MCP servers (Smithery, Anthropic's official registry).
- Hosted runtimes — run remote MCP servers without operating them yourself (Smithery, Composio, mcp.run, MCP Host, Heroku).
- Gateways — single endpoint that aggregates many MCP servers (Composio, Bifrost, Arcade).
- Toolshed-style internal MCPs — your own MCP server fronting your own internal systems (Stripe's Toolshed model — see Approaches: Stripe Minions).
Registries & Marketplaces
| Project | License | What it lists | Notes |
|---|---|---|---|
| Official MCP Registry (registry.modelcontextprotocol.io) | OSS | Canonical community registry | Backed by the official MCP working group |
| Smithery (smithery.ai) | Listing OSS / runtime hosted | 7,000+ servers — installable locally via CLI or run on Smithery's infra as hosted remote servers | "Docker Hub for MCP"; manages runtime + OAuth modals for server authors |
| Anthropic Marketplace (claudemarketplaces.com) | Community | Plugins, skills, MCP servers | 770+ MCP servers as of May 2026; surfaces alongside Claude Code's Discover tab |
| mcp.run (docs.mcp.run) | OSS (Dylibso / Extism) | Wasm-based "servlets" | Each tool is a portable WebAssembly module; install dynamically into a single MCP server, run anywhere |
Hosted MCP Runtimes
Run MCP servers without managing the process yourself.
| Vendor | License | Runtime model | Notes |
|---|---|---|---|
| Composio MCP (composio.dev) | Hosted | Aggregator — single endpoint to a managed library of pre-built integrations | Strongest pre-built tool catalog (Slack, Gmail, Linear, GitHub, etc.); pairs with their agent-orchestrator OSS |
| Coral (withcoral.com) | OSS (Apache 2.0), Rust, local-first | SQL-over-APIs runtime exposing GitHub / Sentry / Datadog / Slack / Linear / Stripe / OTel as queryable tables | Different shape from the rest of this table: instead of one MCP call per action, the agent writes a SQL query that JOINs across sources. Published +31% accuracy / 3.4× cost-efficient vs direct provider MCPs inside Claude Code (Opus 4.6, n=82). See Tool Design § SQL-over-APIs for the architectural argument. ~10 bundled sources at launch — integration breadth is the load-bearing risk |
| Smithery | Hosted runtime | Auto-managed remote MCP servers + OAuth | Same project as the registry — install + run in one click |
| MCP Host (mcp.host) | Hosted | Managed hosting platform for MCP servers | Build-and-deploy without operating infra |
| Heroku AI Apps | Hosted | Heroku dynos for MCP servers | Enterprise-tier managed hosting for production MCP deployments |
| Cloudflare MCP (blog.cloudflare.com) | Cloudflare cloud | Edge-deployed MCP with identity-aware auth | Works with Cloudflare Sandboxes for full agent + tools edge stack |
| Pipedream MCP | Hosted | 2,500+ pre-integrated apps as MCP tools | Workflow-builder roots; broad SaaS coverage |
Enterprise SaaS-integration platforms (now agent-capable)
A separate cluster from the developer-facing MCP servers above — these companies built unified APIs / embedded iPaaS for SaaS-to-SaaS integration first, then pivoted into MCP / agent tool-calling once the protocol won. Sell into the enterprise buyer (compliance, SLAs, per-customer config) rather than the developer buyer.
| Vendor | License | Original product | Agent product |
|---|---|---|---|
| Merge (merge.dev) | Proprietary | Unified API across HRIS / ATS / CRM / accounting / ticketing | Agent Handler (MCP tool-calling) + Gateway (LLM router); positions on enterprise trust + bundled compliance |
| Nango (nango.dev) | OSS (Elastic License 2.0) | Code-first, customizable integration infrastructure | MCP support; positions hard on "AI agents can build and maintain custom connectors" vs Merge's fixed schema |
| Paragon (useparagon.com) | Proprietary | Embedded iPaaS with native integration UI for end users | ActionKit (agent tool platform); often called the best overall Merge alternative for native integrations |
| Apideck (apideck.com) | Proprietary | Unify API across accounting / HRIS / CRM / ATS | Real-time, cache-free unified API positioned for agents; usage-based pricing |
| Scalekit (scalekit.com) | Proprietary | Auth + multi-tenancy for B2B SaaS | Positions explicitly against Merge and Composio on per-user delegation + per-operation scope enforcement for production agents |
The strategic convergence: Merge, Nango, Paragon, Apideck, and Scalekit are all racing toward the same destination (the integration + tool-calling layer for enterprise agents) from different starting points. The wedge each defends is different — Merge on enterprise trust + breadth, Nango on customization + OSS, Paragon on native end-user UX, Apideck on real-time architecture, Scalekit on agent-grade auth. The critique competitors make of Merge specifically: its auth model was built for developer-initiated API calls, while agents need per-user delegation, per-operation scope enforcement, and sub-second execution — an architectural critique worth weighing if you're picking between them.
MCP Gateways
Aggregate dozens of MCP servers behind one OpenAI-compatible (or MCP-compatible) endpoint.
| Vendor | License | Notable for |
|---|---|---|
| Composio Gateway | Hosted | Single unified endpoint for hundreds of integrations |
| Bifrost (maximhq/bifrost) | OSS (Apache 2.0) | Acts as both MCP client and server — unifies LLM gateway + MCP gateway in one Go binary |
| Arcade (arcade.dev) | Hosted | DevOps-leaning — registry + gateway + runtime in one product |
| Obot (obot.ai) | OSS + cloud | Enterprise MCP gateway with role-based access and audit logging |
Choosing
- Single team building one agent — Install local MCP servers from the official registry; no gateway needed.
- Many SaaS integrations needed quickly — Composio MCP or Pipedream MCP for the catalog breadth.
- Multi-team / multi-tenant production — A gateway (Composio Gateway, Bifrost, Arcade, Obot) so you can route, budget, and audit centrally.
- Need code-portable, sandboxed tools — mcp.run's Wasm servlet model.
- Already on Cloudflare / edge-first — Cloudflare MCP with identity-aware auth pairs naturally with their Sandboxes.
Agent Identity, Auth & Secrets
Once agents start hitting real APIs and SaaS systems, "what credentials does the agent run as?" becomes a hard problem. The 2026 stack splits roughly into three layers:
- Identity — proving which agent is making a request (scoped tokens, DIDs, workload identities).
- Credential brokering — agents never see raw secrets; they request short-lived tokens from a vault that stays out-of-band.
- Governance & runtime security — runtime policy enforcement, audit trails, and behavioral constraints on agent actions.
Tooling
| Project | License | Layer | Notes |
|---|---|---|---|
| Microsoft Agent Governance Toolkit (opensource.microsoft.com) | OSS (MIT) | Runtime governance | Cross-language (Python, TypeScript, Rust, Go, .NET); policy enforcement + audit + identity in one toolkit. Released April 2 2026 |
| ZeroID (helpnetsecurity.com) | OSS | Identity | Identity & credentialing for autonomous agents and multi-agent systems; explicitly not repurposed human IAM |
| Agent Vault (infisical.com) | OSS | Credential brokering | Credential proxy: secrets stay in the vault, agent receives short-lived scoped tokens |
| Amazon Bedrock AgentCore Identity (aws.amazon.com) | AWS managed | Identity + brokering | First-party AWS implementation; integrates with IAM, Secrets Manager, and Bedrock-hosted agents |
| Cloudflare Sandbox Auth (blog.cloudflare.com) | Cloudflare cloud | Identity-aware sandbox | Per-sandbox identity, dynamic auth, scoped egress — paired with Cloudflare Sandboxes |
| Kagenti (kagenti.github.io) | OSS | Deploy / govern / secure | K8s-native agent governance — policies and identity for K8s-deployed agents |
| AgentField cryptographic identity | Open Source (Apache 2.0) | Identity | W3C DIDs + verifiable credentials per agent — see AgentField |
| Moltbook (moltbook.com) | Proprietary | Identity + social | "The front page of the agent internet" — agent social network with Moltbook identity used to authenticate to third-party apps; agent ownership verified through X. Early access for developers building agent-auth flows. Speculative / niche but the rare example of consumer-shaped agent identity |
Reference architectures
- Credential brokering pattern — The agent never sees raw API keys. It requests a scoped, short-lived token from a vault (Agent Vault, Bedrock AgentCore Identity, AWS STS-style flow) tied to its workload identity. The vault enforces what the agent is allowed to ask for.
- Multi-agent trust — DIDs (decentralized identifiers, Ed25519) per agent, signed inter-agent messages, and trust scoring (e.g. AgentField's 0–1000 trust score with five behavioral tiers).
- Runtime governance — A policy engine (Microsoft Agent Governance Toolkit, Kagenti) wraps the harness and audits / blocks specific actions. Pairs with Guardrails for content policy.
Endpoint inventory & supply-chain response
The runtime-identity tooling above answers "as whom is the agent running?" — but a parallel question has gotten louder since ClawJacked (CVE-2026-25253) and the wave of malicious MCP servers: which developer machines, right now, have the compromised package or extension installed? SBOMs say what shipped to prod; EDR says what ran; neither captures messy local dev state.
| Project | License | Layer | Notes |
|---|---|---|---|
| Bumblebee (perplexityai/bumblebee) | OSS (Apache 2.0) | Endpoint inventory | Perplexity's read-only supply-chain inventory scanner for developer endpoints. Single static Go binary (Go 1.25+, zero non-stdlib deps), zero package-manager execution. Scans lockfiles, package-manager metadata, extension manifests, and MCP configurations across npm / PyPI / Go / RubyGems / Composer / etc., emits NDJSON, and matches against bundled threat-intel catalogs. Three scan profiles: baseline (globals), project (targeted), deep (broad). 2.1K stars |
The IR workflow: when an advisory drops, run Bumblebee across the fleet, get an exact-match list of compromised machines, then escalate from there. Complements rather than replaces the runtime tools above — it's the between SBOM and EDR layer.
Why this is its own category
The repo's Memory and Guardrails sections cover what an agent knows and what it's allowed to say; identity / auth / secrets covers what it's allowed to do, as whom, with what credentials. The August 2026 EU AI Act high-risk-systems deadline is pulling enterprise interest into this category fast — expect this section to grow.
What Stripe Uses
Stripe's devbox infrastructure is essentially a custom agent sandbox platform built on AWS EC2:
- Pre-warmed EC2 instance pools for ~10 second spin-up
- Full dev environment with source code and services pre-loaded
- Isolated from production and internet
- Identical to human developer machines
- No git worktree overhead — full VM isolation
Sandbox vs. Serverless: Fit-for-Purpose
A critical distinction often missed: code execution sandboxes (E2B, Sprites, Daytona) are designed for short-running, developer-oriented workflows — executing code, running tests, isolating discrete tasks. They are not architected for long-running, always-on agents. Serverless agent-ready platforms (Nebius Serverless, AWS Lambda, Modal) address this gap with faster cold-start wake times, container-based GPU execution, and billing models aligned with agent utilization rather than continuous reservation. Teams should select based on workload duration, not just cost per compute unit.
Key Trends
- Security is the primary buying criterion — The ClawJacked vulnerability (CVE-2026-25253) exposed credential isolation weaknesses across the ecosystem. Demand for dedicated VM/microVM isolation has surged. Only a minority of platforms offer hardware-level isolation guarantees.
- Cost optimization via model routing — LLM API costs comprise 60-80% of agent operating expenses. See Inference for details on intelligent routing strategies.
- Checkpoint/hibernate patterns — Sprites.dev's ~300ms checkpoint/restore and auto-sleep after 30 seconds of idle time represents a shift from always-on to always-available agent infrastructure. Full state preservation with zero idle cost is expected to become standard.
- Multi-agent composition drives orchestration demand — As deployments mature from single agents to multi-agent systems, demand for durable orchestration layers (Temporal, LangGraph) is accelerating.
Choosing Your Stack
Starter Stack (Low investment)
- Inference: Direct API (Anthropic or OpenAI) — see Inference
- Agent: Claude Code, OpenHands, or OpenCode
- Compute: Local Docker or GitHub Actions
- Cost: Pay-per-use API tokens only
Growth Stack (Medium investment)
- Inference: Direct API + LiteLLM gateway for routing and observability
- Agent: OpenHands or Open SWE with custom rule files
- Compute: E2B or Modal for sandboxed execution
- Orchestration: GitHub Actions or Slack bot for triggers
- Cost: ~$0.50-5 per agent run depending on complexity
Scale Stack (Medium-high investment)
- Inference: Nebius Token Factory for self-hosted open models + direct API for proprietary models
- Agent: OpenHands or custom harness with multi-model routing
- Compute: Nebius managed Kubernetes with GPU clusters for model serving + Contree or E2B for agent sandboxes
- Sandbox superpower: Contree's Git-like branching enables tree-of-thought agent workflows — fork at each decision, evaluate in parallel, rollback on failure
- Orchestration: AgentField or Symphony
- Advantage: Best price-performance for teams running open-weight models at high volume. Nebius's serverless inference autoscales with agent demand, and the KV-aware routing keeps latency low during multi-turn agent loops. Single-provider stack (Contree sandboxes + Nebius inference + Nebius GPU clusters) simplifies operations.
- Cost: Predictable token-level pricing, significantly lower than API providers at volume
Enterprise Stack (High investment)
- Inference: Multi-provider with Portkey gateway, tiered model routing. Nebius GPU clusters for self-hosted models, direct API for proprietary.
- Agent: Custom agent harness (like Stripe's Goose fork)
- Compute: Dedicated VMs or K8s cluster with pre-warmed pools (Nebius offers managed K8s with up to thousands of GPUs)
- Orchestration: Custom blueprint/workflow engine
- Context: Centralized MCP server (like Stripe's Toolshed)
- Cost: Significant infrastructure investment, but amortized across 1,000+ agent runs/week
Decision Framework
| Question | If Yes... | If No... |
|---|---|---|
| Need full OS isolation per run? | E2B, Modal, or custom VMs | Docker or git worktrees |
| Running 100+ agents/day? | Dedicated infrastructure, K8s | Serverless (Modal, Lambda) |
| Need macOS-specific capabilities? | Cloud Mac (MacStadium, AWS EC2 Mac, Scaleway) | Linux sandboxes |
| Handling production credentials? | Dedicated VM / microVM (Coral, AgentComputer, Nebius VM) | Shared-kernel OK |
| EU data sovereignty required? | Hetzner, OVHcloud, Contabo, Scaleway, Flow Swiss | Global providers |
| Long-running always-on agents? | Serverless agent-ready (Nebius, Lambda, Modal) | Ephemeral sandboxes |
| Budget-constrained at scale? | Self-hosted (Hetzner, Contabo) with custom hardening | Managed platforms |
See also: Inference Solutions for choosing your LLM provider.