Managed vs self-hosted agent hosting: which model fits you

Before you compare features, you face a more basic choice: does someone run your agents for you, or do you run them yourself? Almost every platform we track lands on one side of this line — or offers both — and the decision shapes cost, control, and how much undifferentiated plumbing you own.

The managed model

A managed platform operates the agent runtime for you: provisioning, scaling, and lifecycle are the vendor's problem. Examples in our catalog include Alfe, Relevance AI, Lindy, and Zapier Agents, plus IDE-style products like Replit Agent. Browse the ranked set at /best/managed-agent-platforms.

Strengths: fastest time-to-value, built-in channels and connectors, no infrastructure to babysit. Trade-offs: less control, potential lock-in, and — a recurring theme — usage-based billing that is harder to forecast. Several managed products meter by credits, "activities," or "actions," and some don't publish exact figures.

The self-hosted model

Self-hosting means you run an open-source framework or SDK on your own infrastructure. The core is usually free, and you own the deployment. Strong open-source options include LangGraph (MIT), CrewAI (MIT), Mastra (Apache-2.0 core), and Cloudflare Agents (MIT, on Cloudflare's runtime). See open-source frameworks.

Strengths: control, portability, no per-seat platform fee, data stays in your environment. Trade-offs: you own scaling, memory, and ops. Read licenses carefully — some tools that look "open" (n8n's Sustainable Use License, AutoGPT's Polyform Shield) restrict commercial use.

Hybrid and infrastructure options

The line isn't always clean:

• Hybrid: many frameworks ship an open-source core plus a managed cloud (LangGraph, CrewAI, Mastra, Dify, n8n) — self-host to start, move to managed at scale. Note that fully self-hosted or hybrid deployment is sometimes gated to enterprise tiers.
• Infrastructure: Modal, Fly.io Machines, and E2B are neither managed agents nor frameworks — they're compute you assemble the agent layer on top of. They lead on scalability.

How to decide

• Optimising for speed and low ops? Go managed — just model the metered cost first.
• Need control, portability, or data residency? Self-host an open-source core.
• Want both? Pick a hybrid framework and keep the exit door open.
• Building bespoke infrastructure? Start from compute (Modal, Fly, E2B) and own the stack.

Key takeaways

• The first decision is who runs the agent — managed vs self-hosted.
• Managed wins on time-to-value; watch for hard-to-forecast, metered billing.
• Self-hosted wins on control and portability; read the license before you assume "open source."
• Hybrid frameworks and raw compute are real middle paths — compare at /platforms.