Container infrastructure, purpose-built for MCP
- Dedicated machines per deployment – no noisy-neighbour hops
- Health checks against /ping with retry-on-failure
- Optional persistent volumes mounted at /data
MCP Server Hosting
Deploy private or public MCP servers on dedicated infrastructure – with a gateway, call logs, per-tool access control, and directory distribution built in.
Your code
Glama · build · host · gateway · logs · analytics
Claude · Cursor · ChatGPT
Every call logged · Every tool gated · Every credential managed
Running an MCP server in production is deceptively hard
You do not just need somewhere to run a process. You need authentication that matches the MCP specification, TLS that terminates correctly for Streamable HTTP, session handling that does not break behind a load balancer, OAuth credential refresh, per-user rate limits, secret rotation, audit logs, and runtime monitoring that understands JSON-RPC. Most teams underestimate every layer.
A 2026 scan of 2,181 remote MCP endpoints found 52% completely dead and only 9% healthy. A separate 2025 scan of internet-exposed MCP servers found 100% missing authentication entirely.
Glama has operated more than 20,000 MCP server instances. Every class of failure a production MCP deployment can hit – protocol quirks, auth drift, infrastructure flakiness – is one we have already engineered our way through. You inherit the scar tissue.
Ship an MCP server, not an ops project.
What you get on day one
Four things every serious MCP deployment needs, shipped together so you do not have to stitch them up yourself.
Gateway in front of every deployment
- OAuth 2.1 credentials managed with automatic refresh
- Per-tool access control – enable or disable individual tools
- Every JSON-RPC call logged with full payloads for audit
Observability built for MCP, not copy-pasted from generic infra
- Runtime logs tailed live from the machine
- Tool-call analytics across 1h, 24h, 7d, 14d, 30d, 90d windows
- Automated health tests that understand the MCP handshake
Secure by default
- Environment variables encrypted at rest
- Access tokens scoped to connection profiles
- Private by default – flip to public listing when ready
Three ways to deploy
Pick the path that matches where your MCP server lives today. The output is the same in all three cases: a Gateway endpoint your clients can point at.
From the registry
Pick any server from the Glama registry and deploy it to your own instance in one click. Useful when you want a private copy of a community server under your control.
- Browse /mcp/servers and pick a server
- Click Deploy on Glama
- Paste the generated Gateway URL into your client
From GitHub
Connect a GitHub repository and Glama builds the image straight from source. Best for MCP servers you are actively developing in-house.
- Install the Glama GitHub App on your organization
- Pick a repository; Glama reads glama.json or your Dockerfile
- Deploy on demand and roll back by redeploying an earlier build
From a package or image
Ship a Dockerfile, an npm package, or a PyPI module. Configure Node or Python versions, build steps, and CMD arguments – or let Glama infer them.
- Point at a Dockerfile, an npm package, or a PyPI module
- Set environment variables (encrypted at rest)
- Glama provisions a machine and issues a Gateway endpoint
https://glama.ai/endpoints/<your-connection-profile>/mcp
Private by default. Public when you’re ready.
Every deployment starts private and access-gated. When you decide to distribute – whether to your company or to the public – flip a single setting and the deployment gets a directory page where people are already searching.
Private deployment
- Endpoint gated by connection-profile access tokens
- Shared with members of your workspace by role
- Not indexed, not discoverable, not listed anywhere public
- Right answer for internal CRM, Jira, or Confluence MCP servers
Public deployment
- Listed in the Glama directory under your account
- Indexed by tool search across every server in the registry
- Shown on category, integration, and client pages
- Right answer for SaaS vendors shipping an MCP server to their users
Directory distribution is unique to Glama. Fly.io, Cloudflare Workers, and self-hosted stacks give you compute; they do not give you a place to be found. Glama’s directory is the single most trafficked index of MCP servers on the web – see the indexing methodology for how the ranking works.
Observability, built for MCP
Three layers of visibility that treat your server as an MCP server – not as an opaque Linux process that happens to listen on TCP.
Runtime logs
Hundreds of the most recent runtime log events are tailed live from the machine. Filter by time window, grep for a string, or scroll to the moment a particular request arrived.
Tool-call analytics
Every tool invocation is recorded as a telemetry event. Break calls down per tool, per user, and per time window – from the last minute up to the last ninety days.
Health tests that understand MCP
Glama runs automated health tests that drive the MCP handshake – not a generic TCP or HTTP probe. When a deployment stops responding, the health flag flips before your users start seeing errors.
How Glama compares
Fly.io and Cloudflare are great general-purpose compute. Glama is shaped specifically for MCP – which means everything below is built in, not bolted on.
| Capability | Glama | Fly.io | Cloudflare Workers | Self-hosted |
|---|---|---|---|---|
| Container hosting | —Workers runtime | you operate | ||
| Docker image support | ||||
| One-click deploy from an MCP registry | ||||
| Gateway included (auth · ACL · call logs) | —separate product | |||
| Managed OAuth 2.1 with refresh | — | |||
| Tool-level analytics | —generic metrics | —generic metrics | ||
| MCP-aware health checks | —generic HTTP | —generic HTTP | ||
| Browser-based inspector / playground | ||||
| Public directory listing with SEO | ||||
| Encrypted environment variables | —DIY |
Hosting questions, answered
Everything teams typically ask before putting an MCP server on Glama.
Is my hosted MCP server private by default?
Yes. Every deployment is private by default – access is scoped to your account and to workspace members you explicitly add. Requests to the endpoint are gated by access tokens issued per connection profile. You can flip a deployment to public at any time and list it in the Glama directory.
What transport does the hosted endpoint use?
The public endpoint speaks Streamable HTTP – the transport the MCP specification standardized in 2025, which every modern MCP client understands. Servers written for stdio are wrapped automatically, so you do not have to rewrite the transport layer yourself.
What runtimes and build formats are supported?
Dockerfile, npm, and PyPI. Pick a base image and Node version, or point Glama at an npm package or PyPI module and we handle the build. Custom build steps and CMD arguments are both configurable per deployment.
Can I deploy from a GitHub repository?
Install the Glama GitHub App and connect the repository that contains your MCP server. Glama pulls the source, builds the image, and provisions a machine. Re-deploys are triggered from the dashboard or via webhook.
How do I connect Claude, Cursor, or ChatGPT to my hosted server?
Paste the Gateway URL into the client’s remote-server configuration. Claude, Cursor, ChatGPT, VS Code, Zed, and every other MCP-compatible client accept a single streamable-http endpoint – no special client setup is required.
How are secrets and access tokens stored?
Environment variables are encrypted at rest with a server-side key; plaintext values are never stored. They are decrypted in memory only at the moment the machine starts or a process reads them. Access tokens are hashed; the fingerprint shown in the UI is derived from the first and last characters of the original token.
How do I give my team access to a hosted server?
Share the deployment with your workspace. Members of the workspace inherit access to the deployment’s endpoint and logs according to their workspace role, so you manage access to an MCP server the same way you manage access to any other workspace resource.
What happens when my server crashes?
Machines run with a retry-on-failure restart policy and an HTTP health check against /ping. If the process keeps failing, the deployment is marked unhealthy and surfaces in the analytics dashboard. You can inspect the last hundreds of runtime log events from the deployment’s logs tab.
What is the difference between hosting and the Gateway?
Hosting runs your code. The Gateway fronts the traffic. Every server you deploy on Glama gets a Gateway endpoint automatically, so the two are the same control plane from the operator’s point of view. See the Gateway overview for what the proxy layer adds on top of raw hosting.
What happens when I make a deployment public?
You keep ownership. A public deployment is listed in the Glama registry under your account, indexed by Glama’s tool search, and shown on category and client pages. You can switch it back to private at any time; users already connected lose access immediately.
How much does hosting cost?
See the pricing page for the current rate card. Hosting is billed on top of your workspace plan; infrastructure cost is metered by machine hours and egress.
Ship your MCP server on infrastructure that’s built for MCP
Plus a Gateway in front of every deployment, a browser-based MCP inspector, and directory distribution when you’re ready to go public.