Introducing Prefect Horizon
Every company will have a context layer.
Some will build it deliberately. Many will build it accidentally. Either way, it's becoming inevitable.
The context layer is where AI agents interface with your business. It's where teams expose proprietary data, tools, and workflows to autonomous systems. It's where you define what agents can see and do, and where you curate and distribute the context that makes automation useful and safe.
We built Prefect Horizon to be the foundation for that layer.
We didn't guess at the requirements. We know them because we're the developers of FastMCP, the standard framework for working with the Model Context Protocol (MCP).
FastMCP was introduced just a few days after the MCP spec was announced, and it was quickly incorporated into the official MCP SDK. Today, it's downloaded a million times a day, and 70% of all MCP servers are powered by some version of it. In fact, we announced FastMCP 3.0 yesterday.
This gave us a front-row seat to the explosive growth of MCP. But it also meant we saw exactly where companies struggle.
MCP is a young technology. While we can smooth out the development experience in FastMCP itself, the sharpest edges for enterprises are deployment and governance.
It is easy to run a server on your laptop. It is incredibly difficult to distribute it to 5,000 colleagues securely. Right now, most enterprises are practicing "vibe governance." They hand agents powerful tools, like the ability to bill a customer, and then they write a polite note asking the agents to use those tools wisely.
You can't prompt-engineer your way out of operational risk.
But every enterprise that's trying to build better guardrails for their agents hits the same walls: Where do I host these servers? How do I know what's been deployed? Who's allowed to access what?
We built Horizon to answer those questions.
The Platform
Horizon gives you the infrastructure to turn MCP from a protocol into a platform.
It's the codification of the best practices we learned building FastMCP, and it's informed by the real failure modes we've already seen at scale.
We've organized the platform into four integrated pillars to handle the full lifecycle of your context layer: Deploy, Registry, Gateway, and Agents.
Horizon Deploy
Horizon Deploy is managed hosting for MCP servers, with the operational basics handled end-to-end: CI/CD integration, scaling, monitoring, branch previews, versioning, and rollbacks. Push code, and Horizon gives you a live, governed endpoint in 60 seconds.
This is also where the "real world" details show up. A surprising amount of MCP work is actually long-running: background tasks, multi-step workflows, subagent delegation, and other operations that shouldn't block an agent in a chat loop. Thanks to FastMCP, Horizon supports the MCP background tasks protocol so agents can kick off work asynchronously and keep moving.
Horizon Registry
FastMCP taught us something simple: MCP doesn't get adopted as a single server. It gets adopted as a long tail.
A GitHub server here. A ticketing server there. An internal billing server. A customer data server. Then ten more. Then someone forks one. Then nobody remembers which version is safe.
Horizon Registry is the place where that sprawl becomes manageable. It's your central catalog of MCP capabilities across first-party servers, third-party servers, and curated endpoints: who owns them, what they expose, and which versions are approved.
It's the answer to: "Where is the GitHub server?" and "Which customer data tool is approved for production?"
Registry is also where curation becomes concrete, because the context era isn't just about collecting tools. It's about shaping decision spaces.
Horizon lets you build remix servers: virtual servers composed from any combination of first-party and third-party capabilities. Pull the exact tools you need from multiple sources and expose them as a single governed endpoint built for a specific workflow. This minimizes both context pollution and the risk of misusing powerful tools.
Remix servers can also be permissioned more strictly than their underlying sources. Your organization can give you broad access and still give your agent narrow access.
That's a critical distinction. The permission model that makes a human effective can make an agent reckless.
Horizon Gateway
If Registry is "what exists," Gateway is "what's allowed."
Every server in Horizon sits behind a gateway. This is the control plane that enterprises actually need: role-based access control down to the tool level, authentication, audit logs, and usage visibility. Horizon Gateway turns MCP tools into governed capabilities.
This is another place where building FastMCP changes the product. When you're at the center of the ecosystem, you learn quickly that "tool access" is not a binary. It's a policy surface. The questions are always the same:
- Who can call this?
- Under what conditions?
- Using which credentials?
- With what audit trail?
Horizon answers those questions at the platform layer, not in prompts and hope.
Horizon Agents
MCP has a last-mile problem: clients.
Even teams that build great servers can fail to deliver value because the client experience is uneven. Managing server connections can be challenging for technical users. For end users, it can be a non-starter.
We know this intimately because we ship the framework. We see where things break. We see where adoption stalls.
Horizon Agents solves the last mile directly.
Horizon includes a fully permissioned interface for chatting with any MCP server or curated combination of servers. You can create agents with a defined universe: the exact context you intended, governed by the Gateway, curated via Remix.
Deploy those agents in Horizon, in Slack, or expose them over MCP for whatever client your company chooses. Business users get an agentic interface to your company without ever needing to learn what "MCP" is.
The Context Era
For almost ten years, Prefect has been obsessed with one thing: giving our users confidence in automated work.
We started by giving data engineers confidence in their scripts. We evolved to give platform teams confidence in their distributed systems. Now, we're giving organizations confidence in their AI agents.
The runtime has changed, from Python scripts to LLMs, but the need for governance, visibility, and trust remains exactly the same. This is automation for the context era.
Welcome to Horizon.
FastMCP Cloud is graduating to become Horizon Deploy, but its most popular feature lives on: free hosting for personal MCP servers. A healthy enterprise ecosystem depends on a thriving open-source community, and we're committed to supporting the developers building the future of MCP.