5 Best MCP Hosting Options for Production GTM Tools

Yananai A. ChiwutaPublished ·13 min readUpdated
5 Best MCP Hosting Options for Production GTM Tools

TL;DR

  • Cloudflare Workers suits lightweight remote servers that fit its runtime and limits; you still need to design access control.
  • Google Cloud Run is a strong option for containerised servers and teams already operating in Google Cloud.
  • Amazon Bedrock AgentCore Runtime supports MCP servers and is worth evaluating when your organisation already builds on AWS.
  • Vercel is a natural candidate for a team with an existing web application and deployment workflow on its platform.
  • Render supports deploying a server as a web service when a conventional application deployment model is preferable.
  • Compare complete operating cost, including identity, secrets, monitoring, external APIs and maintenance. A public endpoint is not a production authorisation design.

What does MCP hosting include?

An MCP server exposes tools or data to a compatible client. Hosting is where that server runs and how the client reaches it. For a GTM team, the server might read account information, look up campaign status or create a task in a CRM. The practical decision is not simply where the code starts: it is who can invoke each tool, how credentials are managed and what happens when a call fails halfway through.

Choose the runtime after identifying your clients and transport. A local process using standard input and output is not automatically a remotely accessible HTTP service. A remote deployment must also handle identity, request limits, logging and the downstream API's permissions. A gateway can sit in front of servers to manage access, but a gateway and the place your code executes are distinct decisions.

The recommendations below are based on provider documentation, not a deployment test. Before committing, run your own server and client combinations against the operational checklist later in this guide. For choosing the tools themselves, see GTM MCP servers; for access across multiple servers, see MCP gateways.


MCP hosting options at a glance

Host Best starting point Key qualification Cost to investigate
Cloudflare Workers Lightweight remote HTTP tools Runtime compatibility, state and authentication design Account plan, requests, CPU and other services
Google Cloud Run Containerised workloads in Google Cloud Container lifecycle, concurrency and client authentication Compute, requests, networking and dependent services
AWS AgentCore Runtime Teams using AWS agent infrastructure Runtime protocol contract, container architecture and session handling Runtime usage, identity, networking and other AWS services
Vercel Web teams deploying MCP alongside an application Function limits, authentication and long-running operations Plan, functions, usage and dependent services
Render Teams that prefer a conventional web service Instance sizing, availability, scaling and auth Service instances, bandwidth and dependent services

These are hosting approaches, not identical packaged MCP products. The right comparison is the cost and responsibility for the same working server under the same access and reliability requirements.


1. Cloudflare Workers: for lightweight remote servers

Cloudflare documents building a remote MCP server on Workers. This can suit an HTTP-facing tool that performs a small amount of processing and calls another service, especially when the team already uses Workers and understands its runtime.

Confirm that your dependencies run in the environment and that CPU, request and subrequest limits fit peak usage. Decide whether the tool is stateless or needs a separate state store. Authentication examples are starting points, not proof that every caller should receive the same downstream access. Map each identity to the permissions it actually needs.

Workers pricing lists a $5 monthly minimum for the paid plan, with included request and CPU allowances. Other services have their own usage rules. That is a useful floor for a model, not an all-in price for a secure production integration.

Choose Workers when the runtime is a fit and the team can implement the surrounding controls. A server requiring an incompatible native dependency or a long-running process may be better suited to a container-based host.


2. Google Cloud Run: for containerised servers

Google's Cloud Run MCP guide describes hosting a server using streamable HTTP and discusses client authentication. A container gives teams a familiar packaging route for a server with application dependencies, while Cloud Run handles the service deployment and scaling model.

The choice still requires decisions about request concurrency, instance limits, timeouts, identity and networking. If a tool calls a private data source, verify that the server can reach it and that the caller is authorised to use the result. If the client runs outside your cloud environment, test the actual authentication flow rather than assuming an internal identity configuration will work unchanged.

Cloud Run is particularly relevant when the team already manages images, IAM and monitoring on Google Cloud. Compare its pricing using your region, configured resources and billing model; add any networking, secrets, databases and external API charges. Serverless scaling does not remove the need to measure cold-start behaviour and downstream rate limits.


3. Amazon Bedrock AgentCore Runtime: for AWS teams

AWS documents deploying MCP servers to AgentCore Runtime, including stateless and stateful streamable-HTTP configurations. Its protocol contract specifies container and session requirements, including ARM64 compatibility and handling the session identifier.

That specificity matters. AWS requires streamable HTTP and recommends stateless mode by default; its stateful mode supports capabilities that need an ongoing session. The server must use ARM64 and listen on 0.0.0.0, with port 8000 by default or the configured serving port. It must accept platform-forwarded session headers. For MCP version 2025-11-25 and earlier, clients capture and reuse Mcp-Session-Id for instance affinity and, with stateful servers, session state. For version 2026-07-28 and later, AWS uses X-Amzn-Bedrock-AgentCore-Runtime-Session-Id, while retaining compatibility with the older header. Use the documented header for the actual protocol version. Test your actual client against the documented invocation and authentication flow, and decide whether stateful features are needed. A tool that calls an AWS-backed data source still needs narrowly scoped permissions; the ability to deploy a server does not grant a safe application-level access policy.

For monitoring, the same AWS contract distinguishes protocol versions. With 2025-11-25 and earlier, or a missing or invalid version header, most runtime errors appear as JSON-RPC errors inside HTTP 200. With 2026-07-28 and later, HTTP status reflects the error. Record both layers and the protocol version: an HTTP-only dashboard can hide failures on older connections, while assumptions about universal HTTP 200 handling are wrong for newer ones.

The protocol version affects that decision. AWS says elicitation and sampling require stateful mode for MCP version 2025-11-25 and earlier, while version 2026-07-28 and later can use multi-round-trip requests without stateful mode. Verify the versions supported by the actual client and server before choosing the session design; do not infer compatibility solely from a host's feature list.

AgentCore is a candidate when AWS is already the organisation's operating environment and its runtime contract matches the server. Model the actual runtime usage and related AWS services using AgentCore pricing. For a simple existing container, also compare other AWS hosting routes on their own terms instead of assuming AgentCore is the only choice.


4. Vercel: for web application teams

Vercel's MCP deployment guide describes using its MCP handler with a deployed application, including an authorisation example. A team already shipping a Next.js application there may value keeping server code, deployment and rollback close to its existing workflow.

Check the function execution model, duration limits, concurrency and dependency support for the exact workload. A CRM lookup that returns quickly has different needs from a research tool that waits on several services. The example authorisation integration still needs an appropriate authorisation server and policies for the resources the tool can reach.

Vercel is worth shortlisting when your web team owns the application and prefers the same delivery workflow. If the server requires a container-specific dependency, long-lived processing or private networking, test those requirements and the applicable plan before choosing. Budget for function usage and connected services rather than treating the existing web subscription as the whole marginal cost.


5. Render: for conventional web service deployment

Render's MCP hosting guide describes deploying a server as a web service with a remote endpoint. This approach can fit teams comfortable with a conventional application service and a source-control deployment process, including teams whose server does not fit a particular edge runtime.

Size the service for memory, expected concurrency and availability. Check instance behaviour when idle, scaling, deployment health checks and the cost of keeping the desired capacity available. An endpoint reachable over HTTPS still needs authentication and authorisation appropriate to its users and tools.

Render is a practical candidate when a straightforward web service is more important than specialised MCP management. Confirm service pricing, bandwidth and any database or identity requirements for your selected plan. If the organisation already has extensive AWS or Google Cloud controls, compare the additional operating environment with using those existing capabilities.


What will three servers cost to operate?

Use the same workload for every quote: three read-only servers, 100,000 tool calls per month in total, and two operating scenarios. In the first, assume ten calls can be in flight simultaneously; in the second, assume 100 can. These are test inputs, not observed traffic. Measure request duration, memory, cold starts and downstream API calls before translating them into provider-specific units. A tool call does not necessarily equal one billable HTTP request.

For illustration, assume a team allocates $40 a month to compute, $25 to logs and storage, $15 to identity and secrets, and four hours of maintenance at $60 an hour. The monthly operating estimate is $320, or $0.0032 per tool call at 100,000 calls. This is a synthetic planning model, not a price for any listed host. It excludes downstream data subscriptions, model calls, setup and incident response. Raising peak concurrency may require more capacity even when monthly calls stay the same; recalculate those inputs after load testing.

To make the peak assumption visible, compare the baseline with a second hypothetical, not yet measured, configuration. Both handle the same 100,000 monthly calls; the second assumes more reserved or provisioned capacity and more operational work for the 100-call peak.

Planning input Ten calls in flight 100 calls in flight
Compute allocation $40 $120
Logs, storage, identity and secrets $40 $40
Maintenance at $60/hour 4 hours = $240 6 hours = $360
Monthly total $320 $520
Cost per 100,000-call workload $0.0032/call $0.0052/call

The extra $200 is a sensitivity example, not a prediction that a tenfold concurrency increase requires triple the compute or two extra staff hours. A provider that scales on demand may have little idle cost but incur different request, duration or cold-start trade-offs. Replace every assumption with the selected host's billing units and the fixture's measured peak behaviour; keep the same availability and latency target in both estimates.

For one bounded provider example, Cloudflare's paid Workers plan lists a $5 monthly account minimum and includes 10 million requests and 30 million CPU milliseconds. If the entire account stays within both allowances, the Workers base component is $5; identity, state, logs and other products may add costs. Do not assume 100,000 tool calls consume exactly 100,000 requests or that all CPU usage fits without measurement. The allowances are shared with other applicable account usage.

As a capacity check, if those 100,000 calls each produced exactly one billable Worker invocation and there were no other account usage, 30 million CPU milliseconds would average 300 milliseconds per call. That is an illustrative division of a shared monthly allowance, not a per-call entitlement or a latency target. Authentication, retries, other Workers and background processing can reduce the available budget; per-invocation limits also still apply.

Ask every provider or internal platform team for an estimate using the same assumptions: three deployments, availability target, average and peak concurrency, compute profile, authentication, log retention and data transfer. Include someone responsible for updating dependencies and responding to an incident. A cheap runtime that no team can maintain is not a complete operating model.


How do you evaluate a host before moving production?

Build one small, read-only fixture: a server that returns a synthetic account record from a controlled dataset. Define two roles, one permitted to read the record and another denied. Use the same server version, expected responses and client for each host. This is a proposed evaluation; we have not deployed or benchmarked the candidates for this article.

  1. Deploy and connect. Record the steps, required account permissions, endpoint, transport and time to reach a working response. Confirm that the actual client supports the chosen authentication flow.
  2. Test isolation. Verify that an unauthorised caller cannot invoke the tool or read another tenant's data. Rotate a credential and confirm the old credential no longer works.
  3. Exercise failure. Simulate a slow or unavailable upstream service. Record timeout behaviour, error messages and whether the client can distinguish failure from an empty result.
  4. Test concurrency and recovery. Try the ten-call and 100-call scenarios, capture latency and errors, then deploy a faulty version and practise rollback. Record cold-start observations rather than accepting a generic claim.
  5. Inspect operations. Trace a request without logging secrets or unnecessary customer data. Confirm who receives alerts, who can change permissions, and how logs are retained and deleted.
  6. Estimate the bill. Use measured requests, duration, CPU, memory and log volume where relevant. Add staff time and the downstream system's costs, then repeat for the busier scenario.

Use an explicit run sheet so the two concurrency scenarios can be compared without confusing a runtime failure with a data failure:

Record for each attempt Why it matters
Host, server version, client and transport An incompatible client or deployment can fail before the tool runs
Start time, cold or warm observation, duration and outcome Separates startup delay from a slow downstream lookup
In-flight calls, rejected calls and retry count Shows whether the 100-call peak causes queuing or repeated work
Expected and returned synthetic account ID A successful response from the wrong tenant is a failed isolation check
Request or session reference and final result Lets the operator distinguish a completed call from one still awaiting reconciliation

Run the same fixture and traffic shape on each candidate, and compare error rate and latency at the same acceptance threshold. Record the region, resource limits and actual billable usage alongside the results. A failed or unfinished call is not an empty account record; retain that status for the client and investigate it before reporting an answer.

The owner of the fixture should also write down how secrets reach the runtime, which deployment can access production credentials, and whether an operator can revoke one user without disabling everyone. Review client compatibility, regional processing and log retention against your own requirements. A successful connection test establishes only that one path worked at that moment; it does not establish an availability guarantee.

For write-capable tools, add idempotency, confirmation and recovery tests before allowing a client to change a CRM. A retry after a timeout must not silently create duplicate tasks. Document ownership and escalation alongside deployment; the GTM engineering playbook helps structure that handoff.


Which MCP hosting option should you choose?

Start with Cloudflare Workers for a lightweight server that fits its runtime, Cloud Run for a container and a Google Cloud operating model, or AgentCore Runtime for a matching AWS workflow. Vercel is worth evaluating alongside an existing web application, while Render is an option for a conventional hosted service. Run the same fixture, identity checks and cost assumptions before selecting a provider.

A note on older shortlists: Smithery says its MCP proxy and managed identity services shut down on 31 October 2026; its public registry remains. Do not treat a directory listing as a currently available proxy or hosting contract. For the larger platform decision, our guide to cloud AI agent platforms covers more than the server runtime alone.


FAQ

Does an MCP server have to be hosted remotely?

No. A client can run a local server over standard input and output. Remote hosting becomes useful when multiple clients or users need a network-accessible service, but it adds responsibilities around identity, availability and operations.

Can I host an MCP server without a gateway?

Yes. A server can expose an appropriately secured remote endpoint without a separate gateway. A gateway may be valuable when an organisation needs a shared policy or access layer across several servers; evaluate that layer separately from the runtime hosting your code.

What should be secured first?

Identify the caller, authorise the specific tool and resource, and protect the downstream credentials. Start with read-only access where possible. Test denial, credential rotation and audit logs using the real clients you intend to support, then add write permissions deliberately.

Yananai A. Chiwuta

Author

Yananai A. Chiwuta

CEO & Co-Founder

Yananai A. Chiwuta is the CEO and Co-Founder of Forma Nôrden, where he builds managed acquisition systems for B2B companies through signal-based outbound and precision paid ad acquisition. He has built and exited two companies, most recently FunnelVision.

Celine Sky-Chiwuta

Article reviewed by

Celine Sky-Chiwuta

Co-Founder & CMO

Celine Sky-Chiwuta is the Co-Founder and CMO of Forma Nôrden, where she shapes the positioning and marketing behind the company’s managed acquisition systems. She previously served as CMO of FunnelVision through its 2025 acquisition.

Related Articles