Host MCP Servers. From git push to live URL.

Connect your GitHub repo and FastMCP Gateway handles the rest: containerization, OAuth 2.1 with dynamic client registration, TLS, autoscaling, and per-request observability. Your server gets a production endpoint that works as a Claude custom connector or with any MCP client.

From the creators of FastMCP

Git-push deploys

Auth handled for you

Scale to zero

Start free

Talk to the experts

Get a custom demo of MCP hosting for your team.

Prefect is trusted in Production.

How It Works

Connect your repo, push, deploy

Connect your GitHub repository to our Gateway. Every push to main triggers a build pipeline: your code is containerized, deployed to an isolated runtime, and assigned a stable URL with OAuth 2.1 already configured. Pull requests get their own preview deployments. No Dockerfiles, no YAML configs, no deployment scripts.

Get started

# Write your MCP server
$ cat server.py
from fastmcp import FastMCP
mcp = FastMCP("my-tools")
@mcp.tool()
def query_database(sql: str) -> dict:
    return db.execute(sql)
# Push to GitHub
$ git push origin main
# Build → containerize → deploy → assign URL
# https://my-tools.horizon.prefect.io

Zero-Config

You write the MCP server. We host the infrastructure.

CI/CD with branch previews

Every commit to main triggers an automatic build and deploy. Open a PR and a preview deployment spins up at its own URL so reviewers can test MCP changes before they reach production. Rollbacks are instant — select any previous deployment and promote it.

OAuth 2.1 with dynamic client registration

Clients that connect to your server negotiate credentials automatically via the OAuth 2.1 dynamic client registration flow — no manual client ID provisioning. This is the auth standard the MCP spec requires for remote servers and that Anthropic mandates for Claude custom connectors.

Autoscaling to zero

Scales up automatically when agents connect and scales to zero when idle. Free hosting for personal MCP servers and pay only for compute you use in production.

MCP observability

Every tool call, resource read, and prompt request is traced with latency, status, and the calling client. Filter by agent, time range, or error type to debug failing requests with full context.

Claude Connectors

Turn your server into a Claude custom connector

Anthropic requires a hosted MCP endpoint that implements OAuth 2.1 with dynamic client registration before Claude can connect. That auth handshake is the part most teams get stuck on.

Our Gateway handles the entire OAuth flow for you: token issuance, client credential negotiation, and token refresh. Deploy your FastMCP server, paste the URL into Claude, and the same endpoint is reusable from any other MCP-compatible AI client.

What the auth flow looks like

Client connects to your server URL and discovers the OAuth metadata automatically.

Dynamic client registration issues credentials on the fly — no manual provisioning step.

Token exchange and refresh happen behind the scenes. Your server code never touches auth.

Background Tasks Protocol

Long-running work without blocking the agent

An agent asks your MCP server to run a data pipeline or kick off a multi-step workflow. Instead of holding the connection open until it times out, the server returns a task handle immediately. The agent keeps working and checks back when results are ready — or gets notified asynchronously.

Our Gateway supports this natively via the MCP background tasks protocol. No polling hacks, no blocked chat loops, no 30-second timeout ceilings.

Built on FastMCP, the open-source framework behind 70% of MCP servers worldwide.

Differentiation

The MCP platform built by the team that built the ecosystem.

No other MCP platform combines open-source framework leadership with full-lifecycle enterprise infrastructure.

Built by the creators of FastMCP

We created the framework behind 70% of MCP servers - we know what breaks at scale.

Full lifecycle, not just one piece

Deploy, catalog, secure, and monitor MCP servers in one platform instead of stitching together point solutions.

Production-grade from day one

OAuth 2.1, RBAC, audit logs, auto-scaling, and monitoring are built in for real production workloads.

Works with the AI clients teams actually use

Any MCP client. Any MCP framework. Claude, Cursor, Windsurf, VS Code, OpenAI, and Bedrock.

Customer Proof

Teams are already running MCP in production with FastMCP Gateway.

Nitorum Capitalcase-study

Horizon is the avenue by which we can best deploy MCP within our organization. Out of the box, take my GitHub repo, launch it, and it just works for us.

James Brink

Head Trader

Theory Venturescase-study

Because we've built these bespoke MCP tools with Horizon, I can stay in my LLM chat UI for a lot longer and get the information I need directly without having to tab through all these different places to copy and paste.

Adam Conway

Senior AI Engineer

Enatacase-study

Horizon actually had the best onboarding experience on my team. Even people who have not done tech before were able to onboard onto the tool, which is a very strong sign of ease of use.

Smit Shah

Founder

Everything you need to run MCP servers in production

Our Gateway includes hosting, auth, and observability on day one, so there is no separate identity service to run and no deploy pipeline to maintain.

Git-push deploys

OAuth 2.1 with dynamic client registration

Branch preview deployments

Autoscaling to zero

Per-request observability

Background tasks protocol

Deploy your first MCP server

Free tier for personal servers. Upgrade when you need team features.

Start free Talk to the experts